Markku Kaskenmaa

USING DATA ANALYTICS IN HOCKEY PLAYER TALENT IDENTIFICATION

Master’s Thesis
USING DATA ANALYTICS IN HOCKEY PLAYER TALENT IDENTIFICATION
Master’s Thesis

Markku Kaskenmaa
Master’s Thesis
Spring 2023
Degree programme in Data Analytics
and Project Management
Oulu University of Applied Sciences
ABSTRACT

Oulu University of Applied Sciences

Degree programme in Data Analytics and Project Management

Author(s): Markku Kaskenmaa

Title of the thesis: Using Data Analytics In Hockey Player Talent Identification
Thesis examiner(s): Dr. Teppo Räisänen
Term and year of thesis completion: Spring 2023 Pages: 90 + 2 appendices

Goal for the research is to identify relevant data and features which can help the team to find rising
future talent to build a data and analytics model and simulate the analysis with real data. The aim
and main objectives are to create both data and statistical models for analysing a hockey game
data in relation to the defined indicators for identifying better than usual player performance and to
run the analysis with real life data.

Data analytics in professional sports can still be seen as a relatively new concept. In professional
hockey leagues, like Liiga or NHL, a lot of data related to games are collected, analysed, and even
made publicly available.

This research is conducted by analysing the previous research documentation, articles, and results,
by collecting information from trusted web sites, blogs, and articles and using them to define the
data, features, and statistical models, and building data model and data pipeline and finally
executing an example analysis with the build models and collected real data. This research has
quantitative and qualitative research parts. Qualitative parts goal is to define the theoretical
framework and the quantitative part includes building a data pipeline and visualizations, and
running the analysis based on the indicators and model built in the qualitative part.

There are numerous statistics and metrics which all try to explain player performance. Data is
collected from games and made available via commercial or public channels, like Wisehockey for
Finnish Elite Hockey League, Liiga. In analysing player performance, focus should be put in what
happens when player is on ice. With analysing data from games and by using machine learning to
detect anomalies in players performance can help teams in player recruitment decisions. Machine
learning based predictions and analysing video data from games can help teams in future not only
on recruitment but forming winning composition and identifying game patterns during game. The
key to realizing these benefits is access to relevant and trustworthy data.

Future research should focus on using novel methods like machine learning to bring full benefits
from data and analytics.

Keywords: Data Analytics, Ice Hockey, Data Analytics in Professional Sports, Machine
Learning, Ice Hockey Statistics

3
CONTENTS

1 INTRODUCTION .................................................................................................................... 7
2 RESEARCH AIM AND OBJECTIVES ..................................................................................... 8
2.1 Aim and objectives ...................................................................................................... 8
2.2 Hypothesis and research questions ............................................................................ 8
3 RESEARCH METHODOLOGY ............................................................................................. 10
4 ORGANIZATION ................................................................................................................... 11
5 THEORY ............................................................................................................................... 12
5.1 Use of data and analytics in professional sports ....................................................... 12
5.2 Use of data and analytics in hockey player recruitment ............................................ 13
5.3 Using game data to analyze hockey player performance .......................................... 14
5.4 Use of machine learning to predict future performance ............................................. 16
5.5 Data analytics and analysis ....................................................................................... 16
5.6 Liiga in general .......................................................................................................... 17
5.7 HockeyAllSvenskan in general .................................................................................. 17
5.8 Traditional hockey statistics, metrics, and their meaning .......................................... 18
5.8.1 Plus/Minus................................................................................................... 18
5.8.2 Time on Ice ................................................................................................. 19
5.8.3 Points .......................................................................................................... 19
5.8.4 Goals ........................................................................................................... 19
5.8.5 Games played ............................................................................................. 19
5.9 Advanced hockey statistics, metrics, and their meaning ........................................... 20
5.9.1 Corsi ............................................................................................................ 20
5.9.2 Fenwick ....................................................................................................... 21
5.9.3 PDO ............................................................................................................ 22
5.9.4 Goals Versus Treshold................................................................................ 23
5.9.5 Point Shares................................................................................................ 24
5.9.6 Expected goals............................................................................................ 26
5.9.7 Zone starts .................................................................................................. 27
5.9.8 Defensive Efficiency Metrics ....................................................................... 27
5.9.9 Game Impact Metric .................................................................................... 28
5.9.10 Players streak duration ............................................................................... 29

4
 5.9.11 Total Hockey Rating .................................................................................... 30
5.9.12 dCorsi .......................................................................................................... 31
5.9.13 Other advanced statistics ............................................................................ 33
5.10 Evaluating players ..................................................................................................... 33
5.10.1 Evaluation of players performance.............................................................. 33
5.10.2 Projecting expected player value ................................................................ 33
5.10.3 Laws of ice hockey analytics ....................................................................... 34
6 DATA ANALYSIS .................................................................................................................. 38
6.1 Data sources ............................................................................................................. 38
6.1.1 Liiga matchlevel data, Wisehockey ............................................................. 38
6.1.2 HockeyAllsvenskan data ............................................................................. 43
6.1.3 Season level data, SportContract ............................................................... 44
6.1.4 Statistical model .......................................................................................... 45
6.1.5 Data model and datasets ............................................................................ 46
6.2 Analytics architecture ................................................................................................ 49
6.2.1 Architecture considerations ......................................................................... 49
6.2.2 Architecture and data processing flow ........................................................ 51
6.2.3 Extracting data from API ............................................................................. 52
6.2.4 Ingesting data to datalake ........................................................................... 53
6.2.5 Storing data ................................................................................................. 54
6.2.6 Processing and transforming ...................................................................... 55
6.2.7 Building queries and combining datasets.................................................... 56
6.2.8 Tables and views for analytics .................................................................... 57
6.2.9 Visualization ................................................................................................ 58
6.2.10 Machine learning ......................................................................................... 58
6.2.11 Security and data protection ....................................................................... 60
6.3 Analytics results ......................................................................................................... 61
6.3.1 Anomaly detection....................................................................................... 61
6.3.2 Normalising and summarising ..................................................................... 63
6.3.3 Expected goals ranking ............................................................................... 63
6.3.4 Analysis against reference group ................................................................ 65
7 CONCLUSION AND FINDINGS ........................................................................................... 75
7.1 Findings ..................................................................................................................... 75
7.2 Hypothesis testing ..................................................................................................... 77

5
 7.3 Other findings ............................................................................................................ 78
8 DISCUSSION ........................................................................................................................ 79
8.1 Challenges, issues and resolutions ........................................................................... 79
8.2 Data and results reliability ......................................................................................... 81
8.3 Future research possibilities ...................................................................................... 81
9 ACKNOWLEDGEMENTS ..................................................................................................... 83
REFERENCES .............................................................................................................................. 84
APPENDICES ............................................................................................................................... 91

6
1 INTRODUCTION

This research is my master’s thesis in OAMK’s master’s degree programme in data analytics and
project management.

The purpose for this research is to help professional hockey team KalPa with their player
recruitment for their Finnish Elite League (Liiga) team. Goal for the research is to identify relevant
data and features which can help the team to find rising future talent to build a data and analytics
model and simulate the analysis with real data. For the sports teams who operate under a moderate
financial budget it is important to identify the future team players as early as possible and with cost
efficiently.

In professional hockey leagues, like Liiga or NHL, a lot of data related to games are collected,
analysed, and even made publicly available. However, the extent to the teams utilizes this data for
the player recruitment process varies. In professional sports like soccer, American football and
baseball, the data is used for many purposes. Within the NHL, the teams are using the collected
data in multiple ways, for example as part of the team coaching and for identifying exceptional
talent.

In this research the focus is on game data which is available through Liiga provided data sources
and HockeyAllsvenskan league data, available either directly from the league database, or via other
trusted public sources. Before starting the research, it has also been identified that in addition to
analysing the game data, another important part of the player recruitment is their personal
characteristics. These are left out by intention, and focus is put solely on the game data to ensure
the research scope is manageable within the scope of the thesis.

This research is conducted by analysing the previous research documentation. These include
articles, and results, by collecting information from trusted web sites, blogs, and articles and using
them to define the data, features, and statistical models, by interviewing team coaches, scout and
building data model and tools using Amazon Web Services and finally executing an example
analysis with the build models and collected real data.

7
2 RESEARCH AIM AND OBJECTIVES

2.1 Aim and objectives

The aim and main objectives are to create both data and statistical models for analysing a hockey
game data in relation to the defined indicators for identifying better than usual player performance.
In addition to the creation of the models, the objective is also to run the analysis with real life data.

Secondary objectives are to identify relevant indicators which can be used to identify better than
usual and exceptional performance of hockey players in their career by analysing the available
game data.

2.2 Hypothesis and research questions

The hypothesis for the research is that by using commonly known ice hockey statistics to analyse
game data and by using machine learning models, a professional hockey team's management,
coaches and scouts can identify and predict players performance and use it to support their player
recruitment decisions.

RQ1: How to identify players whose performance development is better than normal by using data
and analytics?

Goal for this research question is to identify the indicators, statistical model and data which can
identify better than usual and exceptional players (outliers). The players can be young rising talent
up to players who have been playing in the different leagues, like Liiga or HockeyAllsvenskan for
several seasons. Indicators, statistical and data models should not rely only on goals scored and
assist to goal points.

RQ2: What advanced metrics and statistics models are common in hockey, and can they be used
to analyse Liiga and HockeyAllsvenskan game data?

Goal for this research question is to identify what advanced metrics and models are common when
analysing hockey data in professional leagues such as NHL and how well they fit for purpose in

8
analysing Liiga and HockeyAllsvenskan game data? What additional data would be needed to use
these metrics and models?

RQ3: What data is available?

Goal for this research question is to identify what data is available for a Liiga hockey team and what
are the trusted sources to obtain such data? The sources should be such from where the data can
be retrieved constantly and automatically preferably via API and should not be only for one time
purpose (e.g., by not using exported Excel file at time specific moment).

9
3 RESEARCH METHODOLOGY

This research is a mixed method research containing both quantitative and qualitative research
parts. The research starts with qualitative literature-based research, of which the goal is to define
the theoretical framework, to study the relevant previous research and their results to identify the
indicators, and data and statistical models to analyse these indicators for identifying player
performance. The sources for the qualitative part are previous academic research reports, articles,
and results, and trusted publicly available blogs, websites, and articles. In this part I will research
how the data is used for talent identification in other relevant professional sports and leagues in
hockey (e.g., National Hockey League, NHL and Elitserien), soccer (e.g., Premier League),
baseball (e.g., Major League Baseball, MLB) and American football (e.g., National Football
Leagues, NFL) and if these are applicable to national league players in Finland.

The qualitative part can include semi structured interviews which aim is to verify the models fit for
purpose. Interviewees can include coaches, team management, talent scouts, and other teams’
data analysts (for example Nashville Predators data analysts).

The quantitative part includes building a data pipeline and visualizations, collecting the data, and
running the analysis based on the indicators and model built in the qualitative part. Data analysis
results and their relevance are then evaluated, and the model and methods can be adjusted as a
result for future analysis. The data will be obtained from Liiga data sources (Wisehockey) and
HockeyAllsvenskan league data by retrieving it directly from available API’s. The pipeline,
visualizations and analytics will be built using public cloud capabilities and services by using
Microsoft Visual Studio Code with Python 3.9 for coding the data extraction, S3, Athena, Glue and
QuickSight from Amazon Web Services for the analytics and by programming languages commonly
used in data analytics, like Python and SQL.

It is important to note that Finnish national league, Liiga, and Swedish league HockeyAllsvenskan
both collect and provide access to data and statistics from the games. However, the leagues in the
scope for this study may not provide similar data and statistics used in the references in this theory
part, the statistical and data models in this study must meet the data available from selected Liiga
and HockeyAllsvenskan.

10
4 ORGANIZATION

The target organization for this research is KalPa Hockey (KalPa). KalPa is a sports team from
Kuopio, Finland and it was founded in 1929. The team specialised in playing ice hockey in the
1950’s. KalPa has played in the Finnish Elite League, Liiga since 2005 and in history KalPa has
played also in lower divisions. KalPa Hockey Oy’s turnover in 2020 was 6,7 million euros and in
2021 4,6 million euros. (KalPa Hockey)

KalPa is active in providing opportunities for both girls and boys in junior level players. Junior KalPa
“aims also to develop players for the professional team. Its most important goal is to work closely
with the community to make sure local children get a well-rounded start to life”. “Junior KalPa has
teams for every age group, ranging from the youngest juniors up to 20-year-old players. Junior
KalPa also has skating and hockey schools for younger children. Junior KalPa has its own teams
for girls and women. There are over 1,000 players and supervisors in Junior KalPa every year.”
(Junior KalPa)

Thesis researcher is Oulu University of Applied Sciences MBA student Markku Kaskenmaa and
supervisor is Dr. Teppo Räisänen. KalPa supervisor is Anssi Laine.

11
5 THEORY

5.1 Use of data and analytics in professional sports

Today, analytics has become an important tool for any organization to be successful. In
professional sports, analytics is often associated with player recruitment decisions. In a data-driven
world, professional sport teams are using analytics to confirm or predict answers to questions, like
ticket, player recruitment and performance, coaching, pricing and customer relationship. (Mondello,
Kamke, 2014)

It’s not just teams and the league that is using big data analytics. Sport betting companies have
professional hockey as one of the top sports on offer and with in-depth data, they can determine
the bets and winning multipliers more precisely. Game producers and game companies may use
this data to design their sports games. (Julie, 2014)

Hockey and other professional sports are a huge entertainment business. The fans, media and
other analysts can use the data to get and share insights for the player and teams’ performance
and the game details. There is an analytics business around these sports. Companies that provide
data analysis are not just doing this for the fans. They sell their analytics to the teams, players,
coaches, opponents and media and it is their business. Media companies like MTV Media, ESPN
need the analytics that tracking game and player data provides. (Julie, 2014)

In hockey, the puck and players move on the rink with speed that makes manual statistics
impossible to use. Sensors and The Internet of Things have changed data collection for many
businesses, including professional sports. There can be sensors in the puck and all over hockey
players which all can help to gain better performance overview which coaches and scouts can use
to recognize talents under the radar. (Tran, 2021)

Data can be used in many ways in hockey. It can for example indicate fatigue, and coaches can
react instantly by for example changing combinations or rest players in some games. Data analysis
can bring tools for training and schematics for tactics, game plans and team play. In addition, data
can give a more objective view in their development areas for seasoned players and rising young
talent. (Hakes, Sauer 2006)

12
Data analytics in professional sports can still be seen as a relatively new concept. One of the most
known stories is the Moneyball book and movie which is a story of how the Major League Baseball
(MLB) team Oakland Athletics successfully overcome baseball's player market by using statistics
and data. (Hakes, Sauer 2006)

Hakes and Sauer introduced and economic evaluation of the Moneyball story by applying standard
econometric procedures to data on player compensation and productivity. Their findings were that
specific baseball skills were not valued correctly, and that by using statistical knowledge team
manager were successfully able to improve the player valuation. (Hakes, Sauer 2006)

5.2 Use of data and analytics in hockey player recruitment

One of the most important topics in sport analytics is the evaluation of player performance. Analytics
can be used by coaches and team management for coaching, managing team, in scouting players,
and in entertainment. (Vik, Min-Chun, Jansher, Carlsson, Lambrix, 2021)

Even though there have been major development efforts in past years, hockey still lacks many of
the analytical data tools that are in use in other sports like in baseball, basketball, and football.
(Tran, 2021)

Hockey teams can use statistics and data to evaluate their current players and use that knowledge
then to future player decisions. By using analytics for objectively quantifying and analysing their
teams and players contributions can reveal inconsistencies and further give indication where to
allocate player salary budgets. Methods for player valuation can be unique to a sport and are not
consistent across different sports. The primary means in nearly all professional sports for talent
recruitment is through a draft. A draft is a process in which the professional teams take turns in
selecting players from the pools of players. (Brower, 2020)

Talent selection and identification in professional sports is complex, and coaches and teams can
utilize physiological and technical performance assessments to identify future top players.
Identification of these future high valued players can be seen as one of a coach's key tasks and
abilities. Talent identification has traditionally been based on viewing and analysing athletes in trial
games or training sessions. This approach is based on a subjective view of the ideal player, relies

13
on viewers personal opinions and experience, and can include bias and limited consistency.
(Larkin, O’Connor, 2017)

One of the main goals for talent identification and ice hockey player recruitment is to identify and
select the most promising young players with the potential to become a successful professional
player. In ice hockey, there are five main variables which influence what level a player can compete:
technical, tactical, physical, social, and mental. To be a good player one must be faster, stronger,
and smarter. (Kontsas, Lehtola, 2014)

As successful team sport players performance is a combination of multiple skills and qualities, and
team coaches and scouts must consider a holistic approach for talent identification (Larkin,
O’Connor, 2017). Player comparison requires deep domain knowledge including defining
appropriate key metrics for players, and finding a group of players who have playing styles which
can match together and make a winning team. Ice hockey leagues provide in-game data and to be
agile on player markets teams should learn to use a data-driven approach to evaluate and predict
the players performance. (Vik, Min-Chun, Jansher, Carlsson, Lambrix, 2021)

5.3 Using game data to analyze hockey player performance

There are multiple statistical methods that attempt to quantify hockey player’s contribution and
performance in relation to others. Some of these are more traditional metrics such as goals, assists,
points and some are novel and more complex (Pallotta, 2021). These more novel and complex
include for example Goals Versus Threshold developed by Tom Awad, Corsi rating, Behindthenet
Rating by Gabriel Desjardin, Player Contribution by Alan Ryder, and Even-Strength Total Rating
by Timo Seppä (Macdonald, 2011). Other examples of the common metrics are time on ice (TOI),
the games played (GP), Shots on Goal, Expected Goals (Piccolo, 2022), Goals For per 60 and
Goals Against per 60 (Pallotta, 2021).

Corsi is a statistic commonly used in ice hockey measures even-strength play shot attempt
differential (Schulte, Zhao, 2017). NHL.com defines shot attempts as: “any time a player tries to
shoot the puck” and “any shot on goal, blocked shot or missed shot is classified as a shot attempt”
(Piccolo, 2022). Another commonly known metric is Fenwick in which only missed shots or shots
on goal are counted and blocked shots excluded (Piccolo, 2022). Goals Versus Threshold, GVT,

14
is a statistic which tries to quantify the overall contribution to the game of a given player. This
statistic is a combination of Offensive, Defensive, Goalkeeper and Shootout GVTs (Tran, 2021).

Such metrics have strengths and weaknesses. One weakness is that many of the statistics exclude
the information about game context where it was measured and assigned (Schulte, Zhao, 2017).
Another weakness was concluded by Schuckersa and Argeris in their research paper: “teams were
not substantially better at picking optimal players with respect to TOI, GP, or GVT but were
markedly better at selecting players within a half standard deviation of optimal”. (Tran, 2021)

To evaluate a players performance, it is more common to use traditional metrics that are related to
players the actions, like using a sum of scored goals goal and assists. Some of the statistics used
to evaluate a players performance try to consider the game context, like how the event impacted
to the game result. A scored goal, when the team is in the clear multiple goal lead at the end of the
game does not define the winner of the game. But if a player scores a goal only some seconds
before end of the game in a situation where game is tied, many times defines winner. Some players
tend to score these important goals than others, and other player may score more often these not
so important goals. In example, Washington Capitals player Alex Ovechkin was ranked highest is
making these important goals in the NHL season of 2013-2014. In scoring not so important goals
his rank was 29th. The game context where the goal is scored should be considered when
evaluating players' performance. (Vik, Min-Chun, Jansher, Carlsson, Lambrix, 2021)

Multiple studies exist in the context of using analytics in player evaluation in professional sports.
Players are often rated by their observed performance over a series of games. There is an
increasing trend for using data analytics and machine learning to model ice hockey. Teams, team
management, coaches and fans are interested to evaluate players. For example, for the team
management, questions like which players to sign, trade or draft are common. Multiple studies
about different models exists, and the most common model is to quantify the total value of players
based on their actions. Traditional sports analysis models usually and exclude the actions that lead
up to the goal and their focus is on the actions that immediately impact to the scoring goals, as an
example passes and pass reception. Values provided by these methods are limited. The relevant
context may include for example time on ice, player role or position, current score, and game
situation. (Liu, Schulte, 2018)

15
5.4 Use of machine learning to predict future performance

There are studies for novel methods like machine learning algorithms and models for player
evaluation. Liu and Schulte in their study article “Deep Reinforcement Learning in Ice Hockey for
Context-Aware Player Evaluation” introduced a new method for evaluating game context by using
Deep Reinforcement Learning (DRL). In their method they firs use learn an action-value Q function
from game events. Then by using a puck possession based Long Short-Term Memory, the neural
network integrates continuous history and game signals. Their method then uses the learned Q-
function to valuate players actions also considering game contexts and evaluate a player
performance. They call the valuation, which aggregates players actions and contribution, as Goal
Impact Metric (GIM). They also found that a players GIM stays consistent throughout a season and
has high correlation with future player salary and success measures. (Liu, Schulte, 2018)

5.5 Data analytics and analysis

To define the term “data analytics” in the context of this research its common definition will be used.
Cambridge Dictionary specifies data analysis as “the process of examining information, especially
using a computer, in order to find something out, or to help with making decisions.” (Cambridge
Dictionary)

Data analysis is a process of exploring, refining, transforming, and training models with the chosen
relevant data to gain useful information for decision making. Data analytics uses methods and tools
like machine learning, statistical analysis, and pattern detection to gain insights from data. Its goal
is to help organisations in decision making and problem solving by using historical data. Data
analysis is often seen as a subset of data analytics. (Sarangam, 2022)

Google defines data analysis as a six phased process. The phases are:
• Ask. This first phase is about asking the right questions and understanding why analysis
is done and what kind of problem it tries to solve.
• Prepare. The second phase is to prepare the data for analysis. This means collecting
relevant data to the problem and to define metrics that are needed for the analysis.
• Process. This phase is about processing the data. Data is checked and corrected against
various inaccuracies, errors, and inconsistencies so that it is consistent and is credible and
trustworthy for the analysis.

16
 • Analyse. In this phase data is analysed. Primary goals for this phase are finding the
relationships, trends, and patterns that will help to solve the problem. This phase is about
finding what the data is telling.
• Share. The fifth phase is about sharing findings. This can be done with help of visualization,
storytelling and using narratives.
• Act. In this sixth phase, all things learned from analysis is put in action by providing
recommendations to the stakeholder on problem solving and helping them make decisions.
(Gautam, 2021)

5.6 Liiga in general

Finnish Elite League, The Liiga (formerly SM-Liiga), is the top professional ice hockey league in
Finland. Liiga was founded in 1975 to replace the fundamentally amateur league SM-sarja. Liiga is
and Finnish Ice Hockey Association have a cooperation agreement, but Liiga is not directly
controlled by latter. (Wikipedia)

In a Liiga season consists of the regular season and playoffs there are currently 15 teams playing.
All these teams play 60 matches during a regular season, a quadruple round robin with extra local
double rounds. (Wikipedia)

The six highest scored teams per points of the regular season proceed directly to quarterfinals.
Teams ranked between seven and ten play preliminary play-offs with best out of three games and
the winners continue to quarterfinals. Quarterfinal winners meet at semi-finals and losers of the
semi-finals play a bronze medal match and the two winner teams play a gold medal match. In the
regular season 2021-2022 there were 632 players in total playing in Liiga. (Wikipedia)

5.7 HockeyAllSvenskan in general

HockeyAllsvenskan is a second highest professional league (previously Allsvenskan and

SuperAllsvenskan) in the Swedish ice hockey system (after the SHL). Since the 2009–10 season,
the league has consisted of 14 teams. In a regular season, all teams play 52 games, and all teams
play each other four times during one season, two home games and two visiting other team.
Following the regular season Slutspelserie tournament is played which finals winner wins the
championships (Wikipedia)

17
In the regular season 2021-2022 there were more than 650 players in total playing in
HockeyAllsvenskan. (Wikipedia)

5.8 Traditional hockey statistics, metrics, and their meaning

This chapter introduces some known traditional statistics and methods which can be used to
explain and evaluate ice hockey players performance.

5.8.1 Plus/Minus

Plus/minus purpose is to measures what effect player has on goals of his team when he is on ice.
When plus/minus value is positive for a player it means that the player has been on ice for more
goals scored than against, and if it plus/minus is negative, it means player has been on ice for more
against. Plus/minus calculates all even strength and short-handed goals, including those where
goalkeeper has been pulled from the rink. (Kohl, 2016)

Plus/minus calculation is based on:

• A plus point is given when a player is on the ice and his team scores.
• A minus point is given when a player is on the ice and opponent scores a goal.
• No points are given when a power play goal is scored.
• Players who are on the ice for an empty net goal will receive a plus and a minus unless
the team who scores the empty net goal is on a power play. (Jones)

Plus/Minus can be seen as misleading statistics if used alone for evaluating players. This is
because the individual player has only little control over their own plus-minus. Players can make
mistakes and they can also excel. Ice hockey is a game which includes randomness. This can
mean that goals are scored, and a player could have had nothing to do with it, and a player is still
given a plus or minus point. Another example of lack of context plus/minus doesn’t consider
goalkeeper skills. Additionally, there is variation in how many goals a team scores during a season.
This means that one player's plus/minus does not necessarily compare to another team's player
and that plus/minus requires normalization or to be used in conjunction with other metrics.
(Thomas, 2018)

18
Plus/Minus is available through both data sources and can be used in the analysis part of this
thesis.

5.8.2 Time on Ice

Time on ice (TOI), or ice time, is a metric that indicates a player's total ice time over the season or
in a single game. For player performance analytics purposes, it is mainly used for normalizing other
metrics. (Sans, Carlsson, Lambrix, 2019)

Statistic is available through both data sources and can be used in the analysis part of this thesis.

5.8.3 Points

Points is a common metric summarizing simply goal leading assists and goals a player has
performed in a single game, over the season or over his career. (Sans, Carlsson, Lambrix, 2019)

Statistic is available through both data sources and can be used in the analysis part of this thesis.

5.8.4 Goals

Goals is a common metric summarizing goals a player made, in a single game, over the season or
over his career. (Sans, Carlsson, Lambrix, 2019)

Statistic is available through both data sources and can be used in the analysis part of this thesis.

5.8.5 Games played

Games played, GP, is a common statistic collected in hockey and it summarizes the number of
games where a player has been on ice over a time. Time is often calculated as a single season or
over a player's career. (Sans, Carlsson, Lambrix, 2019)

This statistic is available through both data sources and can be used in the analysis part of this
thesis.

19
5.9 Advanced hockey statistics, metrics, and their meaning

This chapter introduces some known advanced statistics and methods which try to explain and
evaluate ice hockey players performance. It should be noted that many of the advanced statistics
are intended to be used with the data from the National Hockey League, NHL.

5.9.1 Corsi

Corsi, named after Jim Corsi, a former professional goalkeeper, can be seen as one of the
foundational concepts in hockey statistics. Corsi is based on shot attempts, and considers any shot
attempts, not only shots on goal. This means that any shot which are directed towards the goal is
counted, whether it reaches the goalkeeper or not, or is blocked. A positive Corsi for a team
indicates they took more shots than the opponent team, which can indicate more offensive pressure
and possession. When Corsi is reported on a player it simply indicates how many shot attempts
were made when a given player was on ice and can indicate players contribution to the game result.
(Lee, 2020)

The Corsi formulas are as follows:

𝐶𝑜𝑟𝑠𝑖 (𝐹𝑜𝑟 𝑜𝑟 𝐴𝑔𝑎𝑖𝑛𝑠𝑡) = Shot Attempts For or Shot Attempts Against

Corsi For (CF)

𝐶𝑜𝑟𝑠𝑖 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑎𝑔𝑒 (𝐶%) =
𝐶𝑜𝑟𝑠𝑖 𝐴𝑔𝑎𝑖𝑛𝑠𝑡 (𝐶𝐴)

As Corsi considers all shot attempts, Corsi For percentage (CF%), which averages over total shot
attempts, can indicate a player's performance. CF% is calculated as
CF
𝐶𝐹% =
𝐶𝐹 + 𝐶𝐴

If players CF% is around 50% this can indicate that a player is effective on creating offensive threat
to the opponent but not very active on defence. CF% more than 50 can indicate that a player
contributes equally to offense and defence, or is dominant in one, and good enough in the other. A
relative metric, Corsi For %, is needed to compare a player to a reference. A higher CF% may can
identify more offensive pressure and can lead to more goals. (Lee, 2020) In Corsi, this is CF%
relative and is calculated as

20
 CF%
𝐶𝐹% 𝑟𝑒𝑙 =
𝐶𝐹𝑜𝑓𝑓%

where CFoff% means the team metric where a player was not on ice. CF% rel can identify important
players on a team, or a player to pair with a good scoring player. (Lee, 2020)

Corsi compares only even-strength play situations and does not consider game context, situation,
nor the shot attempt quality. Even very low-quality shots are calculated. An underachieving player’s
Corsi may be affected by better performing players who are on ice with that player. CF% used with
average point total could still indicate if a given player is creating opportunities for his team. (Lee,
2020)

Corsi is available through both data sources and can be used in the analysis part of this thesis.

5.9.2 Fenwick

Fenwick also counts any shot that is made on goal during in even-strength game situation. The
difference to Corsi is, that it excludes blocked shots. The formula to calculate Fenwick is similar to
Corsi. (Lee, 2020)

When measuring Fenwick for a team’s it is often presented as a percentage of the total Fenwick of
the game. If a Fenwick For % (FF%) is above 50% it indicates that the team made more unblocked
shot attempts against goal than their opponent team. Fenwick, when applied to individual players,
compares FF% on the ice and FF% when the player is off the ice. A player who is active on offense
and is effective in defence is indicated by higher FF% and can indicate a player contributes more
when he is on the ice and thus is a potentially high performing player. (Lee, 2020)

Evenly distributed blocked shots between opposing teams is seen as equal Corsi For % and
Fenwick For %. Potential issue with Fenwick is that some players tend to block more shots than
others. With such players, this can be seen as significant differences in their CF% and FF%. This
is related to a player’s role; some are playing in defensive roles and as result are blocking more
shots. Fenwick counts any shot that make it through the defence and differs from shots on goal
(SOG), which counts only shots that make their way to the goalie. Another problem with Fenwick

21
is that a blocked shot could still be a result of a goal opportunity, and it excludes the total the
contribution of a player. (Lee, 2020)

Fenwick is available through both data sources and can be used in the analysis part of this thesis.

5.9.3 PDO

PDO is not an acronym. Luck has an important role on the results in ice hockey and PDO can be
said to be a statistic that measures luck. Luck in this context means game results that are outside
normal boundaries and variance of the player's normal performance. PDO can be calculated both
on a team level and for individual players. (Lee, 2021)

A player cannot maintain a long period of shot percentage that is multiple standard deviations
higher or lower than the mean nor any goalkeeper can stop every shot. A better or lower
performance for a player can be also referred to as luck. Due to randomness of the game and low
number of goals, ice hockey can be said to be more affected by luck than some other sports.
Situations caused by randomness during a game can led to a goal and cause the weaker team to
win a game. (Lee, 2021)

PDO is the sum of a team's shooting percentage (S%) and its save percentage (Sv%) during even
strength play. The combined PDO of all teams in the league is 100%. PDO can be used to evaluate
if a team should expect a regression or improvement (Wikipedia). Luckier teams PDO is higher
than 100% and a PDO less than 100% means that the team has been unlucky or has been
performing below its expected skill level. Over the long term, the luck can be estimated to approach
the average (Weissbock, Herna, Inkpen). PDOs tend to approach 100% is often discussed as
PDOs regression. The mean is 100% because with PDO any shot on goal result is either a goal or
a save. Therefore, a games S% and Sv% will always be 100. Summing PDO over a season will
also give 100%. (Lee, 2021)

PDO is often normalized by multiplying it by 10. PDO formulas are:

S%
𝑃𝐷𝑂 =
𝑆𝑣%
where
Goals
𝑆% =
𝑆ℎ𝑜𝑡𝑠 𝑜𝑛 𝐺𝑜𝑎𝑙

22
and
Saves
𝑆𝑣% =
𝑆ℎ𝑜𝑡𝑠 𝑜𝑛 𝐺𝑜𝑎𝑙
When evaluating an individual player, PDO can be an important measure as it indicates what
happens when a given player is on the ice. (Lee, 2021)

PDO is available through Wisehockey API and can be used in the analysis part of this thesis.

5.9.4 Goals Versus Treshold

Goals Versus Threshold (GVT) is a statistic which uses goals added and aims to evaluate a player's
total contribution to the game. GVT is a combination of Offensive and Defensive GVTs. (Turtoro,
2014)

GVT considers both assists and goals. The threshold is a number calculated compared to a
threshold player. For example, a player like Oliver Kapanen leaves KalPa for the rest of the season.
The questions then are for example who is going to have to replace him on the team? What is the
difference between having Kapanen on the team and not having Kapanen on the team? How much
can the replacement players cover losing Kapanen? When a team considers whether to replace a
player, there are at least two ways for replacement: 1) replacing the player from inside own team
or 2) contracting a player outside the team for example other same league teams or from another
league (bgred105, 2012). GVT aims to quantify answers to these questions and to understand the
value of a player like Kapanen or any other player’s GVT. (CLVNNG, 2017)

GVT was created to 1) Compare various players in the NHL, independent of position and 2) provide
a way to quantify the importance of defence. (bgred105, 2012)

Calculating GVT uses complex formulas and requires estimation of assists value versus goals
value. Calculating GVT requires data how much a player has been on ice and identifying and
calculating a threshold value for possible threshold players. GVT is considering both offensive and
defensive metrics. (Turtoro, 2014)

23
According to Turtoro, “The offensive GVT calculates the total value of a player's points and
subtracts the expected point value of a threshold player resulting in the total goals added by the
player.” (Turtoro, 2014)

The defensive GVT is a sum of 1) Goals prevented by preventing shots, 2) plus/minus adjusted to
team and position, and 3) position adjusted players contribution to the goalkeeper GVT. Defensive
GVT shows the prevented expected goals. The total GVT is calculated by adding Offensive GVT
and Defensive GVT together and adding goalkeeper GVT and shootout GVT to the equation.
(CLVNNG, 2017)

According to Tom Awad, the five key characteristics of GVT are that it
1. is measured in goals.
2. compares all players over any period.
3. uses only directly goal leading statistics.
4. has built-in accounting.
5. normalises for the strength of the league. (Yost, 2012)

Since GVT is meant to calculate statistics for the NHL players and require specific adjusted
parameters, it is not directly applicable for Liiga on HockeyAllSvenskan players. And, that it is more
to explain a replacement value of a player than identifying rising talent. It would however be
interesting to study this more in the context of Liiga, to try to find out if and with what changes it
applies. GVT is not discussed further in this thesis.

5.9.5 Point Shares

Point shares is a statistic originating based on Win Shares in baseball by Bill James and was
developed by sports analyst Justin Kubatko in 2011. Aim for Win Shares is to estimate the number
of wins a player has been able to create and it originates from basketball. Kubatko decided to apply
the same idea to ice hockey after he found the statistic to be working for baseball and basketball.
Point share tries to explain players performance at a time in each context. ( (Kubatko, 2011)

In Point Shares, marginal goals for (MGF) and marginal goals against (MGA) are linked to teams’
points. Team level MGF and MGA are calculated with:

24
 7
𝑀𝐺𝐹 = Team Goals × × 𝑇𝑒𝑎𝑚 𝐺𝑎𝑚𝑒𝑠 × 𝐿𝑒𝑎𝑔𝑢𝑒 𝐺𝑜𝑎𝑙𝑠 𝑃𝑒𝑟 𝐺𝑎𝑚𝑒
12
7
𝑀𝐺𝐴 = Z1 + [ × 𝑇𝑒𝑎𝑚 𝐺𝑎𝑚𝑒𝑠 × 𝐿𝑒𝑎𝑔𝑢𝑒 𝐺𝑜𝑎𝑙𝑠 𝑃𝑒𝑟 𝐺𝑎𝑚𝑒
12
− 𝑇𝑒𝑎𝑚 𝐺𝑜𝑎𝑙𝑠 𝐴𝑔𝑎𝑖𝑛𝑠𝑡

At even strength, a team has five skaters and one goalkeeper on ice at a time, six in total.
Each of these players contribute to offense and defence. Skaters contribute five parts to the offense
and five parts to the defence and goalkeepers both parts will contribute to defence. This gives us
in total 12 parts, and 7 of them are defensive. Thus 7/12. (Kubatko, 2011)

Expected points are then calculated from marginal goals for and marginal goals against with:

𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑃𝑜𝑖𝑛𝑡𝑠 = League Points Per Goals × (MGF + MGA)

Offensive Point Shares are then assigned to players based on goals and with time on ice or games
played. Defensive Point Shares are calculated with:

Marginal Goals Against

𝐷𝑒𝑓𝑒𝑛𝑠𝑖𝑣𝑒 𝑃𝑜𝑖𝑛𝑡 𝑆ℎ𝑎𝑟𝑒𝑠 =
𝑀𝑎𝑟𝑔𝑖𝑛𝑎𝑙 𝐺𝑜𝑎𝑙𝑠 𝑃𝑒𝑟 𝑃𝑜𝑖𝑛𝑡

Goalkeeper gets 2/7 of the team's defensive point shares. Number 2/7 comes from that there are
in total 7 defensive points with 6 players who are on ice at a time and goalkeeper gets two of them.
(Kubatko, 2011)

By summing Offensive Point Shares, Defensive Point Shares, and Goalie Point Shares we get the
final Point Shares for each player (Kubatko, 2011). This statistic could be suitable for Liiga and
HockeyAllSvenskan players, but requires some calculations with data, and is dependent on certain
features on the data. As also this statistic is by default part of the datasets offered by neither of the
sources I use for this thesis, it is not discussed during the analysis phase. Calculating this statistic
would require extra coding and calculations, and as its value is unclear for Liiga and
HockeyAllSvenskan, it is not discussed further in this thesis.

25
5.9.6 Expected goals

Low scoring rate makes analysing team and player performance in ice hockey difficult. This is
particularly present when less than a full season's data is used for the analysis. In addition to low
scoring rates, the randomness makes evaluating and predicting performance of teams and players
harder. (Lukas, 2021)

Expected Goal (xG) is a measure which considers that each shot has a like hood of becoming a
goal and not all shots are equal. (Liu, Schulte, 2018). It uses multiple different factors, like shooting
position, and game situation, and calculates mathematically the like hood of a shot attempt
becoming a goal. xG can be presented as a percentage or directly representing a value of the like
hood. xG can also be seen as a measure of shot quality and can present total shot quality a player
makes, thus can indicate players contribution to the game. (MacDonald, 2012.)

There are many uses for an expected goals statistic, and it can be used to evaluate players. As an
example, expected goals may be used to evaluate offensive and defensive performance of an ice
hockey team over a time, or it can be used to evaluate a player's contribution to a single game.
(MacDonald, 2012.)

Expected Goals For (xGF) can be used to calculate what is the shot quality of a players team and
Expected Goals Against (xGA) what is the opposite teams shot quality when a specific player is on
the ice. When xG is normalized with time on ice, it tries to explain the players contribution when he
is on ice (Lukas, 2021). Another way to use xGF and xGA is to evaluate the overall shot quality
percentage a player has generated. An xGF% higher than 50% can indicate that a player is
contributing to create an offensive advantage for their team. (Lukas, 2021)

Expected goals were also highlighted as an important metric by the data analyst of Nashville
Predators. Wisehockey calculates expected goals value for each shot recorded and this statistic
can be obtained via API thus can be used in the analysis phase of this thesis. However, this statistic
is not available from SportContract API.

26
5.9.7 Zone starts

Zone starts is the number of faceoffs in each end of the rink. It considers the total number of
faceoffs, but not the number of faceoffs wins. (Lee, 2021)

Offensive zone faceoff happens when for example the opponents goalkeeper covers the puck, or
the opponent gets a penalty. In theory, zone starts can be a measure of offensive pressure and
can also indicate player deployment and performance. The problem decreasing the value of this
statistic is that an offensive, or defensive faceoff may be awarded to a player even if he had no part
how the puck got into the zone. (Lee, 2021)

Christian Lee in his Hockey Stats article wrote that his analysis showed that during NHL season
2019-2020 the distribution of player zone starts is left-skewed with a mean just below 50 and that
most players had an offensive zone starts between 35-63%. His analysis revealed a positive
correlation between higher offensive zone starts and higher points per game averages. This
suggests that more offensive faceoff wins lead to more pressure finally leading to more shots and
scoring opportunities. He also concluded that the relationship is much weaker with defensemen
because their primary role is preventing the goals and not scoring. (Lee, 2021)

When used as single measure to evaluate players performance, zone starts is not very informative
and it should be incorporated with additional variables, like faceoff wins and other factors that can
indicate if an offensive faceoff can result as a goal (Lee, 2021). Zone starts is not available from
neither of the API’s I used as data sources, thus will not be discussed further in this thesis.

5.9.8 Defensive Efficiency Metrics

Many of the available statistics and methods try to explain offensive performance, but a question
remains that how to measure defensive performance of a team and a player? To answer to
requirement, Raber and Eisenberg introduced a set of efficiency metrics called Defensive Efficiency
Metrics (DEMs) that aims to draw a comprehensive picture of a team’s performance. (Raber,
Eisenberg)
DEMs aim is to indicate how efficient a given team is at defending against following five dangerous
possession-types:

1. Defensive zone start

27
 2. Controlled zone entries
3. Transition chance
4. East-West slot pas, and
5. Penalties.

Raber’s and Eisenberg’s research indicates that if a team is allowing 20 transition chances in 100
possessions, they have an 80% transition prevention score. If the team allows 5 goals on those 20
possessions, they have a 25% transition defence score. Their method used decision tree analysis
to find the relationships between the five possession types and found out that and that relation with
each of the five possession types were statistically significant and that there is a high correlation
with preventing goals. (Raber, Eisenberg)

DEMs is a team level measure and can provide insight into team defensive performance, strengths,
and weaknesses. Decision makers can use DEMs to identify which situations a team struggles in
defending and use that information in recruiting players that can positively contribute to those.
(Raber, Eisenberg)

DEMs, however, are targeted to highlight teams’ defensive capabilities, thus are not very effective
in predicting a player's career or performance. DEM is not available via the data sources used in
this thesis thus is not discussed further.

5.9.9 Game Impact Metric

Game Impact Metric, GIM, were introduced by Guiliang Liu and Oliver Schulte in their research
paper “Deep Reinforcement Learning in Ice Hockey for Context-Aware Player Evaluation” in 2018.
They introduced a new approach for capturing game context with applying Deep Reinforcement
Learning (DRL) to learn an action-value Q function from game events. Their method uses a neural
network representation which integrates both continuous context signals and game history, using
a possession-based Long Short-Term Memory, LSTM. Then the method uses the learned Q-
function to valuate players’ actions under different game contexts and to evaluate a player’s overall
performance. (Liu, Schulte, 2018)

Their method introduces a Goal Impact Metric (GIM) which aggregates the values of the player’s
actions, reflecting other actions than goals. GIM is possession based instead of goal based. They

28
also found out that GIM can be used to also identify undervalued players, GIM is consistent
throughout a season, and that it has high correlation with success measures and future player
salary. (Liu, Schulte, 2018)

The method uses a computer vision provided dataset which provides very detailed level
information, such as coordinates, about game events and player actions over the entire season
(Liu, Schulte, 2018). This may limit its usage on Liiga and HockeyAllsvenskan and GIM is not
available in either of the data sources, and calculation would require a lot of effort. However, this
method supports my findings later in this thesis about using machine learning for predictions and
detecting anomalies.

5.9.10 Players streak duration

In ice hockey, a player's performance varies over the season, and between seasons. Periods where
in consecutive games a player makes goals or assists and is assigned points can be called streaks.
When streaks are used as a metric, it can indicate players that are having better or less successful
periods. Aa a reference to identify a streak (against his own team), for example points from past
five games can be used. A streak is a sequence of games over series of games where the player
plays during a season, and where the player plays better than usually (hot streak) or below
expectations (cold streak). (Fuentes, Carlsson, Lambrix, 2019)

As discussed earlier, in ice hockey goals are rare which means also that long point streaks are
rare. Streaks as a metric can only indicate offensive capabilities of a player. To overcome these
problems, two other metrics can be used: 1) direct impact, and 2) on-ice impact. (Fuentes,
Carlsson, Lambrix, 2019)

Fuentes, Carlsson, and Lambrix present in their research paper: “Player impact measures for
scoring in ice hockey” that “While the on-ice metrics results in slightly longer streaks, all four curves
show clear straight-line behaviour on lin-log scale suggesting that hot-streak durations when using
an individual threshold is exponentially distributed. This itself suggests that hot streaks, when
assessed relative to the players’ average performance over a season, may be memoryless and
recent performance history (including longer streaks) may not add value compared to just reporting
the average performance over the entire season.” (Fuentes, Carlsson, Lambrix, 2019)

29
This means that a hot streak can primarily indicate a good player at a given time but does not
indicate how a player will perform over a season. Fuentes, Carlsson, Lambrix research notes that
good players are more likely to have longer hot streaks, and that long hot streaks are more frequent
with high salaried players. (Fuentes, Carlsson, Lambrix, 2019)

As a conclusion, player hot-streak duration can help predict the performance in the short term, but
not over the season or multiple seasons. Streak duration can be calculated, but as it is not readily
available in the datasets. This however is an interesting metric for machine learning based anomaly
detection, in other words finding hot and cold streaks in players time series data.

5.9.11 Total Hockey Rating

Total Hockey Ratings (THoR) is a statistic to be used for Forwards and Defensemen in NHL. It was
created by Michael Schuckers and Jim Curro in 2013.

THoR uses a statistical model to evaluate players' overall contribution while on the ice. It considers
players all on-ice actions as well as quality of their teammates (QoT), their opponents quality (QoC),
the current score of the game (Score Effects), and where player starts their shift (Zone Starts).
Each game event gets assigned a value which indicates a change that it leads to a goal.
(Schuckers, Curro, 2013)

Basis for THoR is data from NHL Real Time Scoring System (RTSS). RTSS records all events that
occur in every NHL game, and it uses the Play by Play (PBP) files from every game to obtain the
on-ice action events:
• faceoff (FAC)
• hit (HIT)
• giveaway (GIVE)
• takeaway (TAKE)
• blocked shot (BLOCK)
• missed shot (MISS)
• shot on goal (SHOT)
• goal (GOAL)
• penalty (PENL)

30
With using RTSS, the location for each event is known. Zone information (offensive, neutral,
defensive) where the event occurred, is used for HIT, FAC, TAKE, GIVE, MISS, BLOCK or PENL
events and the x and y coordinates for SHOT and GOAL events. (Schuckers, Curro, 2013)

THoR aims to create a rating system for NHL forwards and defensemen by valuating them with
both scoring goals and preventing goals. THoR uses a data model described above and is then
fitted using ridge regression. A high THoR means a player is contributing more to the events that
lead to goals. THoR handles the issue in evaluation when the performance change if a player
changes a team. (Schuckers, Curro, 2013)

THoR’s advantages include that it evaluates players for every event that happens when a player is
on the ice, and it aims to highlight good performance on both offense and defence. To overcome
the randomness of a game, the model uses two regular season’s data. It adjusts for the quality of
other players on the ice and values each event for the impact it has on scoring a goal. (Schuckers,
Curro, 2013)

In their research paper Schuckers and Curro calculated THoR for the top 15 forwards and the top
15 defensemen. Results of their research indicate that THoR gives a complete evaluation of a given
player. (Schuckers, Curro, 2013)

THoR however is again one advanced metric aimed to evaluate NHL players and requires data
from the NHL RTSS system. Thus, its usability to evaluate Liiga and HockeyAllsvenskan player is
limited. While THoR materials and formulas seem to be still available, it is not widely used.

5.9.12 dCorsi

A study made by Stephen Burst indicated that Corsi For and Corsi Against have a weak correlation
with Pearson's R correlation value of -0.13, with a coefficient of determination of 0.02. As a result,
he developed in 2014 Delta Corsi, in short, dCorsi which is the differential between a skater's
expected and observed Corsi values and it aims to evaluate both offensive and defensive
performance for a player (Burtch, 2014). Corsi was discussed earlier in this thesis.

31
In theory, teams that maintain a greater share of puck possession, are more likely to make goals
and as a result win games. dCorsi uses expected Corsi For and expected Corsi Against to evaluate
how a player impacts on shot metrics. dCorsi is the residual differential of the expected Corsi For
and actual Corsi For and the expected Corsi Against and actual Corsi Against. (Burtch, 2014).

dCorsi uses then a multi-variate linear regression to predict what Corsi results are expected for a
given player based on their usage. However, an individual player has only limited control on their
time on and because of this the actual usage is determined through weighting several factors that
may impact players Corsi results. The factors include for example age, position, team, time on ice,
teammates Corsi For and Corsi Against, offensive and defensive faceoffs and faceoff win
percentage.

The model then calculates delta Corsi For per 20 minutes (dCF20) and delta Corsi Against per 20
minutes (dCA20). Once expected CF and expected CA are calculated the two values can be
compared to the players’ observed Corsi results to obtain dCorsi For and dCorsi Against. Finally,
to calculate a player’s dCorsi over a given season, the expected Corsi is determined by summing
dCorsi For and dCorsi Against. (Burtch, 2014)

Values of dCorsi are nearly normally distributed, which means most players dCorsi values are close
to zero. The better and worse player can be identified by looking at the players who are at least
one sigma away from the centre. (Burtch, 2014)

Stephen Burtch verified the developed metric with selecting 100 player randomly over 7-years
population of NHL players. His studies shows that the model is effective and that there is a strong
correlation between expected Corsi and actual measured Corsi (r = 0.7854 and r2 = 0.6168).
(Burtch, 2014)

dCorsi is again one of the statistics which is complex to calculate and requires specific datasets.
dCorsi is not available through the datasets used for this thesis, nor the dataset containing all
needed parameters to calculate it. For this reason, it is not further discussed in this thesis.

32
5.9.13 Other advanced statistics

In addition to the above statistics, with searching web with Google searches that there are multiple
additional statistics which all try to explain players and/or teams’ performance and value. Examples
of these are Wins-Above-Replacement (WAR) and Regularized Adjusted Plus/Minus, RAPM. Both
being targeted to evaluate NHL players and require NHL RTSS data, or similar dataset. Thus, they
cannot directly be applied to Liiga or HockeyAllSvenskan players and are not discussed in the
thesis.

5.10 Evaluating players

5.10.1 Evaluation of players performance

Rob Found in his 2016 research paper suggests that an optimal strategy for evaluating players
performance may be a combination of a player’s ability to produce high quality shots and a player’s
net production of goals. He compares multiple methods to analyse and predict success and his
results suggest that when comparing approaches to evaluating team success, and individual
contributions to that success, no single metric exists. Goals come from shots, shots can only follow
puck possession, and possession is a result of winning faceoffs, getting power plays, and forcing
turnovers. (Found, 2016)

Many of the metrics are evaluating team success. For forward players an approach could be to
predict team and individual success with goal-based metrics. Neither goal nor shot based metrics
are alone effective in evaluating a defenseman. For players in defence roles, statistics such as
defensive Efficiency Metric which consider all possessions, resulting in larger samples, can be
more effective. (Raber, Eisenberg)

5.10.2 Projecting expected player value

The central element of a hockey team building is to try to measure every player's value with a single
number taking into consideration the player salary budget and try to project it to the several seasons
ahead. (Vollman, Fyffe, Awad, 2016)

33
The previously discussed statistics measure may be used to evaluate players' value based on how
good the player was in the history, but to predict players future success and value to his team, a
model which converts these statistics into players expected future performance in current season
and in players remaining career is needed. (Vollman, Fyffe, Awad, 2016)

To predict a future performance of a player, a three-step method, The Marcel Method introduced
by hockey analyst Tom Tango in 2005 can be used:

1. Use weighted average of player’s three most recent seasons.

2. Use a regression of a players performance against league mean in played games.
3. Adjust against age to differentiate between developing and seasoned players. (Vollman,
Fyffe, Awad, 2016)

Random variance in hockey game analytics can be estimated by using for example a simple
statistics method by dividing the available data into two and checking how the halves correlate with
each other. (Vollman, Fyffe, Awad, 2016)

With this method,

• if there is no correlation, then event is completely random.
• if the correlation is weak, assume that the future performance is closer to the league
average than to the actual observed results.
• if the correlation is strong, assume that the actual performance is closer to the observed
values than league average.
• perfect correlation means there is no random variance and observed results can be used
without any regression. (Vollman, Fyffe, Awad, 2016)

The possibility to use machine learning for predictions exists. This could be for example applied so
that several past seasons data is used to train a model and use the current season to create
predictions. This approach is discussed later in this thesis.

5.10.3 Laws of ice hockey analytics

In hockey analytics as well as in any business analytics it important to consider various aspects
before making any analysis or decisions based on the results. Alan Ryder wrote his article “The

34
Ten Laws of Hockey Analytics” in 2008. According to his article the “laws” of ice hockey analytics
are: (Ryder, 2008)

1. Winning a game matter.

Winning games is all that matters. (Ryder, 2008)

2. Goals for and against affect winning.

A team that scores the most goals wins in a single game wins the game. The team which has
a positive average goal differential over the season tends to win more games than it loses.
(Ryder, 2008)

3. Randomness is everywhere.

Statical proof exist that randomness is everywhere and goals in ice hockey occur randomly. It
means that outcomes do not occur pure by luck, they are uncertain and are influenced of
multiple factors such as skill, strategy, and execution. (Ryder, 2008)

4. Winning has a nearly linear relationship to goal differential.

There is a very high correlation between goal differentials and winning the game. Teams that
are underperform in a prediction based on goal differential are better and the ones that
outperform are worse. Thus, simplified model for winning is:

1 (𝑔𝑜𝑎𝑙𝑠 𝑓𝑜𝑟 − 𝑔𝑜𝑎𝑙𝑠 𝑎𝑔𝑎𝑖𝑛𝑠𝑡)

𝑊𝑖𝑛𝑛𝑖𝑛𝑔 % = +
2 (𝑡𝑜𝑡𝑎𝑙 𝑔𝑜𝑎𝑙𝑠 𝑝𝑒𝑟 𝑔𝑎𝑚𝑒 × 𝑔𝑎𝑚𝑒𝑠 𝑝𝑙𝑎𝑦𝑒𝑑)

This can be called “linearity law”. It means that game is the sum of its parts, both scored goals,
saved goals impact the results and the team’s performance is a sum of its individual players'
performance. (Ryder, 2008)

5. Sample size matters

When evaluating players using data, the relation in the data amounts is important to
understand. If a team plays 82 games in an average NHL season and wins 41 of those games,
participates in about 450 goals, 5000 shot-on-goals and 9000 shot-at-goals there is 95%
confidence that the team is neither better than 60% or worse than 40%. The statistical
information from a sample is proportionate to the square root of the sample size. For one game

35
outcome there are 2.3 goals, 7.8 shot-on-goals and 10.6 shot-at-goal events; information from
eight games of shot-at-goal events is equal to 35 games goals or a full season of wins and
losses. In one season a goalkeeper can face 2000 shots and a good forward can make 300
shots. In general, there is more data about goalkeeping than shooting goals. (Ryder, 2008)

6. Expect mean reversion.

Research exists which show that a good luck is indicated by PDO a lot higher than 1 can and
bad luck with PDO much less than 1. PDO values are almost normally distributed, and it is
one of the easiest predictions in hockey is to find PDO outliers. (Ryder, 2008)

7. Respect the data.

Correct data must be used to answer to the questions. The only unambiguous event in a hockey
game is a goal. Shot information may contain errors or bias and ice time may be inaccurate or
wrong if recorded by humans. Random errors in the data can average out as measurement
noise. Systematic errors or biases should be identified and corrected before data is used in
analysis. (Ryder, 2008)

8. Hockey is a team sport.

Team performance is a sum of Individual players performance and there is a much greater
variance in individual performance than in team performance. From the team perspective, only
thing that matters are the shots that ends as goal. This suggests looking towards statistics like
plus/minus. (Ryder, 2008)

9. Puck possession matters.

Analytic efforts should be focused on the shot data. Previous analysis has shown that historic
shot at goal data predicts future winning better than historic winning or goal scoring. (Ryder,
2008)

10. Context matters.

In general, it can be assumed that a player who is on more ice does more. Players' time on ice
on a game should be adjusted to the game time context (per 60 minutes). Blocking shot
attempts during penalty killing and in power play situation getting shot opportunities are the key
tasks. The roles of players vary, and the data for these roles look different. Some players move
puck on the ice, and some does the shooting goals. A metrics of wingman of a high scoring

36
 player different than a wingman of a different role. Generally, making goals against highly
skilled team is more difficult than against lower skill level team, and the metrics from these
games look different. (Ryder, 2008)

These rules can be concluded:

• When evaluating player performance using game data, the player's role should be
considered.
• Only thing that matter in winning are the goals.
• Puck possession can indicate goal opportunities.
• To predict the future, a large enough data set is needed.
• Understanding the data is crucial for the quality of the analysis.
• Identifying outliers is important. (Ryder, 2008)

37
6 DATA ANALYSIS

6.1 Data sources

I used two data sources to obtain match level and season level data: 1) Wisehockey for Liiga game
level data and 2) SportContract for HockeyAllsvenskan data and full season statistics for Liiga.

6.1.1 Liiga matchlevel data, Wisehockey

As discussed earlier in this thesis, there are numerous metrics and methods which all try to indicate
players' performance and can contribute to the projection and prediction of the future. As it is
important to select the most valuable methods and metrics, it is important to look at the actual data
that can be used for analysis. Even how sophisticated the analysis model and method are, it is only
possible to use that if the data available supports that model.

Some leagues like NHL are more open to share their data but Liiga has chosen commercial
Wisehockey with closed API as their data platform. Wisehockey has a nice web-based user
interface where a user can see the data, replay game situation videos etc. Wisehockey has an API
for reading the stored data programmatically, but that is not publicly open. (Wisehockey)

Wisehockey is a commercial professional sports data and analytics platform developed and
operated by a Bitwise Oy. It uses multiple sources for collecting data from a hockey match, including
sensors on puck and players, and video data. Wisehockey offers a web interface for accessing
statistics and video data. Liiga.fi public website provides some basic statistics based on
Wisehockey data. (Bitwise Oy) Wisehockey web interface provides an easy yet limited view to the
stored data and provides some analytics capabilities. Users of the system are not able to create
custom views or queries to the data. This is where the API enters play, however the statistics
available through API are limited. (Bitwise Oy)

In addition to web interface, Wisehockey offers API for accessing its data programmatically.
(Bitwise Oy). However, the data regarding players performance which can be retrieved from API is
limited to basic statistics. Raw data from matches are available but advanced statistics which are
viewable through web interface are not available through API. API provides details about every

38
shot, shift, skaters’ etc. The advanced statistics, such as expected goals, xG, are not available over
API which limits datasets usability and creates overhead on calculating such statistics. This means
that to use those advanced stats, 1) they must be calculated from the raw data or 2) use a data
scraper to read the web interface and extract these statistics from scraped raw data. For sake of
keeping the scope of the thesis in control, I had to choose option 1 and focus only on those basic
statistics, export the expected goals for each shot in each match, and calculate an expected goals
for each player by extracting each shot xG and identifying the shooter and finally summing expected
goals for a player in one match.

Wisehockey API structure is somewhat complicated. Some statistics which are calculated from raw
data are published through it, some are not.

Wisehockey provides ready to use analytics and statistics on players, both skaters and
goalkeepers. Wisehockey divides statistics into categories which are: traditional and team play
statistics, passing and shooting statistics, skating and puck control statistics, ice time and defence
statistics and goalkeeper’s statistics. Please see tables 1 - 4 for details about statistics which are
available in Wisehockey portal, https://hub.wisehockey.com.

Statistic Description
+ Plus. A plus is received by the scoring team’s players on the ice when an even-strength goal or
shorthanded goal is scored.
- Minus. A minus is received by the players of the allowing team on the ice when an even-strength goal
or shorthanded goal is scored.
+/- Plus/Minus. Total sum of plus and minus
A Assists. Number of assists received from shooting, passing or deflecting the puck towards the scoring
teammate or touching it in any other way which enabled a goal
CA 5v5 Corsi Against. Number of shot attempts against at even strength. Shots + blocks + misses.
CF 5v5 Corsi For. Number of shot attempts for at even strength. Shots + blocks + misses.
CF% 5v5 Corsi For percentage. CF / (CF + CA)
CF% 5v5 relative Corsi For percentage . CF% - CFoff% (CFoff% = team’s CF% when a player is off-ice)
Rel
FF% 5v5 Fenwick For percentage. (FF) / (FF + FA) (FF = shots on goal for + missed shots on goal for)
FF% Rel 5v5 relative Fenwick For percentage. FF% - FFoff% (FFoff% = team’s FF% when a player is off-ice)
G Goals. Number of goals scored by a player
oiSH% 5v5 team on-ice shot percentage while a player was on the ice
oiSV% 5v5 team on-ice save percentage while a player was on the ice
P Points. Number of points received by a player for goals scored or assists earned
PDO 5v5 team on-ice shot percentage + save percentage
PIM Penalty minutes a player received in total
PME/min Player momentum effect per minute. How much momentum score was affected on average when a
player was on the ice.
Screens Screening the opponent’s goalkeeper. How many times a player screened a shot
Table1. Traditional and team play (Wisehockey API documentation)

39
 FP DIST Forward pass distance. Distance the puck traveled toward the opponent’s end of the rink by the
player’s passes (m)
Pass% Percentage of successful passes per a player’s all passes
Passes Successful passes per a player’s all passes
Rcvd Received passes. Number of passes received by the player.
Passes
S Number of a player’s shots
S% Percentage of successful shots per a player’s all shots
Screens Number of the times a player screened the opponent’s goalkeeper
S TOP Player’s hardest shot
TP DIST Total pass distance. Distance the puck traveled on the ice by the player’s passes (m)
xGA The sum of the opposing team’s expected goals from the total time the player has been on the ice.
xGavg Average xG value for a player’s shot
xGF xG For. The sum of the whole team’s expected goals from the total time the player has been on the
ice.
xGF% xG For percentage. The relation between a team’s and the opponent’s xG from the total time the
player has been on the ice. >50% means that the team creates better xG than the opponent when
the player is on the ice.
xGF% Relative xG For percentage. The relation between a team’s xG from the total time the player has
Rel been on the ice and xG from the time the player has not been on the ice. A positive value means that
the team creates better xG when the player is on the ice.
xGsum Sum of a player’s all shot xG values
Table 2. Passing, shooting and xG (Wisehockey API documentation)

1st First puck controls in offenses. How many times a player has been the first one to gain control of
Controls the puck in an offense.
AVG Shift Player’s average shift duration (min)
BKS Number of shots blocked by a player
FOW Number of a player’s faceoff wins
FOW% Faceoff win percentage (per a player’s all faceoffs)
PCW Number of a player’s puck contest wins
PCW% Puck contest win percentage (per a player’s all puck contests)
S Shifts Number of a player’s shifts
TOI Player’s time on the ice in total (min)
Table 3. Ice time and defence (Wisehockey API documentation)

GA Number of goals against

GAA Goals against average. Number of goals allowed by a goalkeeper per 60 minutes played
Pass% Percentage of successful passes per a goalkeeper’s all passes
Passes Successful passes per a goalkeeper’s all passes
SA Number of shots against (goals + saves)
SA/min Average number of shots against per 60 seconds
SA Lateral Shots against made from lateral passes
Screened GA Number of screened goals against
Screened S Number of screened shots (goals + saves + blocked + missed)
SV% Goalkeeper’s save percentage
SV Lateral Saves from shots made from lateral passes
xGa xG (expected goals) against
xG Diff xGa minus goals against
xG Diff Lateral xGa from lateral pass shots minus goals allowed from lateral pass shots
xG Diff Screened xGa from screened shots minus goals allowed from screened shots
Table 4. Goalkeeper statistics (Wisehockey API documentation)

40
Wisehockey API provides the following statistics for skaters in every game. Similar summary
statistics are available for the full season. Wisehockey uses the term tournament here. Statistics
for players are (Wisehockey API documentation):
• Player’s top speed
• Time on ice
• Travelled distance
• Average speed while controlling the puck
• Number of accelerations
• Number of decelerations
• Number of successful passes by the player
• The total number of passes by the player
• Number of successfully received passes
• Total distance of the passes by the player
• Total distance of forward passes by the player
• +/-
• Puck control duration and distance
• Forward puck control distance
• Number of puck contest wins and losses
• Speed zone statistics: player’s time spent on different skating speeds
• Number of offensive screens
• Number of blocked shots
• 5v5 Corsi For, 5v5 Corsi Against, 5v5 Corsi For percentage, 5v5 relative Corsi For
percentage
• 5v5 Fenwick For percentage, 5v5 relative Fenwick For percentage
• PDO
• 5v5 team on-ice shot and save percentages while the skater was on ice
For goalkeepers, the statistics are (Wisehockey API documentation):
• Saves
• Goals against
• Saves with screen
• Goals against with screen
• Shots against with screen
• Shots against from lateral pass

41
 • Shot area statistics: the following statistics calculated for various shot areas (at goal, close,
edges, far):
o Saves
o Goals against
o Saves with screen
o Goals against with screen
o Shots against with screen
• Number of successful passes by the goalkeeper
• The total number of passes by the goalkeeper
• Number of successfully received passes
• Total distance of the passes by the goalkeeper
• Total distance of forward passes by the goalkeeper
• Goalkeeper ”shifts”: times when goalkeeper has entered and exited the rink

From the list above it can be observed that for example the expected goals-based statistics for
players are not available at the time of extracting the data via API. However, Wisehockey provides
detailed level additional statistics for the game, which could be used to calculate such metrics.
These detailed statistics include:
• Skater shifts
• Faceoffs
• Offenses
• Puck control
• Shots
• Goals
• Passes
• Blue line crossings (Wisehockey API documentation)

Shot stats include the expected goal value for all shots, which I will use later in the analysis. The
statistic contains following features for all any recorded shot (Wisehockey API documentation):
• Period number and time of the shot
• The shooting team
• The start position of the shot
• The speed of the shot
• The result of the shot (goal, missed, saved, blocked)

42
 • The shooter, blocker, and saver
• Screening players
• The shot area (at goal, close, edges, far) where the shot started
• Whether the shot was made directly after a lateral pass
• The direction of the shot at the goal
• Whether the royal road was crossed before the shot or not
• Team strength information during the shot
• Skating speed of the shooter during the shot
• Expected goals value of the shot

Expected goals (xG) was earlier highlighted as one of the valuable statistics when analysing players
performance. If expected goals is going to be used, it requires to calculate it using raw data by
summarising the xG values in a game for any player for any shot.

Same analogy goes with goals statistics. This creates more complexity to the analysis, requires
more coding, and as a result limits the scope of the analysis phase for this thesis. These statistics
however are available through the web portal, and the recommendation for Wisehockey is to offer
the same statistics for Liiga teams to use via API. Another way to solve this would be using data
scraping with the web portal. The limitations of scraping were discussed earlier in this thesis.

Exporting, cleaning, extracting, combining, and summing the datasets was a heavier task than
expected during planning of this thesis work and required additional coding and steps during data
transformation. Recommendation for the future development and implementation is to use effort to
code scripts to export advanced statistics from the web interface to avoid heavy transformation and
allow us to use the already existing advanced statistics.

6.1.2 HockeyAllsvenskan data

I found out that HockeyAllSvenskan data can be obtained either directly from HockeyAllSvenskan
website or via analytics websites such as SportContract. However, the amount of data points from
the games compared to Liiga is less advanced.

43
6.1.3 Season level data, SportContract

One of the goals for this thesis was to extract and ingest the data programmatically via API.
SportContract is yet another commercial professional sports data, video, and analytics platform.
SportContract invents and develops new ways to improve professional sports analytics and
scouting, built on top of the latest technologies. SportContract does not tell in their public websites
their data sources. One possible option is that they read the data from other free and commercial
services. They also provide some data via their API, and for the purpose of the thesis I had access
to both SportContract web interface and to their API. (SportContract).

Similar problems with statistics are present with SportContract API; many of the statistics shown
in their web portal are not programmatically accessible through API calls. Specifically, advanced
statistics like expected goals are missing. SportContract API however provides the same features
for both Liiga and HockeyAllSvenskan over full regular season 2021-2022, so the full season
dataset is obtained via their API. Please see table 5 for a skater full season dataset details
(SportContract API documentation).

id Players id in the system

firstName Players first name
lastName Players last name
position Players role
teamId Players teams id in the system
teamName Players teams name
jerseyNumber Jersey number
GP Games played
G Goals scored
A Assists
P Points
PM Penalty Minutes
PlusMinus PlusMinus points
PP Power Play time
SH Short-handed time
AS All shots
TOIPerGame Time on ice per game
FOW Won Faceoffs
FOL Lost Faceoffs
FO_perc Faceoffs win percentage
BLS Block shots
SOG Shot on goals
Hits Hits
A1 Primary Assist
A2 Secondary Assist
PPG Powerplay Goals
SHG Short-handed Goals
Table 5. SportContract full season dataset.

44
SportContract has both a web interface and API. Their web-interface has advanced and traditional
statistics, but their API is limited to basic statistics over the season. Limited amount of game data
can be read per transaction (10 games) and the game statistics are limited. This would have meant
that the queries would become complex and the value of that data low. At the time of extracting
data, per game stats API was not available. (SportContract API documentation)

Full season statistics are easily accessible over API, and I used both Liiga and HockeyAllSvenskan
full season datasets exported from SportContract API. Neither SportContract makes advanced
statistics such as expected goals available via their API. (SportContract API documentation)

6.1.4 Statistical model

As shown earlier in this thesis, here are various statistics and models which all try to evaluate and
predict player performance. Many of these are developed for NHL, and because of that use of those
models would require NHL specific data, such as the RTSS data. Many of them are not generally
in use in Finnish and Swedish leagues and their capabilities or feasibility for this purpose cannot
be verified in the scope of this thesis.

I interviewed one of the NHL data analysts during my thesis work and based on this interview the
most crucial is to select metrics and statistics which indicate how the team is doing when the
specific player is on ice, meaning what is their contribution to the team and to the result of a game.
This view supports my findings on the researched advanced statistics.

Because of the additional coding that would be needed for the data transformation and to calculate
the advanced statistics I needed to limit the scope of the analysis phase.
To keep the scope of my thesis manageable, I limit the statistics used to analyse Liiga players
performance-based match data on Wisehockey dataset to the following:

• Time on ice - To be used to normalize the other features. Tells how much time a player is
playing.
• Goals - Measures players' contribution to winning the games.
• PDO - Measures “luck” and randomness.
• Expected goals For - Measures the offensive pressure a player has when on ice.
• Plus/Minus - Measures a players contribution to winning the game.

45
 • CorsiFor and Corsi% - Measures how many shot attempts were made when a given player
was on ice and can indicate players contribution to the game result.

Based on the theory studied in the thesis earlier, this set of statistics should give a reasonable view
on a player's performance using historical data.

6.1.5 Data model and datasets

For this thesis, and to keep the thesis scope manageable, I decided to narrow down the datasets
to be used in example analysis in one full regular season. This further means that I decided to use
a full season data from the Finnish Elite League, Liiga, from the most recent regular season period,
which is the 2021-2022 season. Full season consists of 450 matches in total, with 15 teams playing
meaning 60 matches per team (Wikipedia). Full regular season offers variance in games, all teams
play against, whereas in playoffs and finals, only certain teams play against each other, which could
of course affect the results of the analysis. I used the same analogy for HockeyAllSvenskan data,
the most recent data for full season is regular season 2021-2022. In this league, in the regular
season, there were 14 teams playing, 52 games per each team, and in total 364 games (Wikipedia).

From the API sources I extracted the following datasets. All datasets contain regular season 2021-
2022 data. As a general comment on the datasets, the source from Wisehockey uses feature
named id for many of different occasions and to avoid mixing them during analysis, there was a
need to rename these features when creating actual analysis queries, tables, and views. I needed
also to add additional identifiers to the datasets, like matchid. Prefiltering, for example dropping
several features not necessary for my thesis, of the raw data and concatenating the results to single
files per dataset was done for easier handling and uploading to the analytics pipeline. Datasets are:

1. Teams
• Source: Wisehockey API
• Contents: All teams played in Liiga
• Features: id, fullName, shortName
2. AllPlayers
• Source: Wisehockey API
• Contents: All players played in Liiga
• Features: id, firstName, lastName, jersey, role, teamId

46
 • Notes: If a player has been playing in more than one team during the season, there
are separate entries for each. This created some challenges during analysis.
3. AllMatches
• Source: Wisehockey API
• Contents: Information about all matches
• Features: id, date, status, homeGoals, awayGoals,
lastUpdatedhomeTeam.id,homeTeam.fullName,
homeTeam.shortNameawayTeam.id, awayTeam.fullName,
awayTeam.shortName, venue.name, venue.city
4. AllMatchStatistics
• Source: Wisehockey API
• Contents: Detailed match statistics for each match
• Features: player, team, totalStatistics.skatingStatistics.topSpeed,
totalStatistics.skatingStatistics.timeOnIce,
totalStatistics.skatingStatistics.distanceTravelled,
totalStatistics.skatingStatistics.averageSpeedWithPuck,
totalStatistics.skatingStatistics.accelerations,
totalStatistics.skatingStatistics.decelerations,
totalStatistics.shiftStatisticsSummary.shifts,
totalStatistics.shiftStatisticsSummary.averageShiftDuration,
totalStatistics.passStatistics.successfulPasses,
totalStatistics.passStatistics.allPasses,
totalStatistics.passStatistics.receivedPasses,
totalStatistics.passStatistics.totalPassDistance,
totalStatistics.passStatistics.forwardPassDistance,
totalStatistics.plusMinusStatistics.plus,
totalStatistics.plusMinusStatistics.minus,
totalStatistics.plusMinusStatistics.total,
totalStatistics.puckControlStatistics.puckControlTime,
totalStatistics.puckControlStatistics.puckControlDistance,
totalStatistics.puckControlStatistics.puckControlForwardDistance,
totalStatistics.puckContestStatistics.puckContestsWon,
totalStatistics.puckContestStatistics.puckContestsLost,
totalStatistics.puckContestStatistics.puckContestParticipations,

47
 totalStatistics.shotScreenStatistics.offensiveScreens,
totalStatistics.shotScreenStatistics.blockedShots,
totalStatistics.traditionalStatistics.corsiFor,
totalStatistics.traditionalStatistics.corsiAgainst,
totalStatistics.traditionalStatistics.corsiForPercentage,
totalStatistics.traditionalStatistics.relativeCorsiForPercentage
totalStatistics.traditionalStatistics.fenwickForPercentage,
totalStatistics.traditionalStatistics.relativeFenwickForPercentage
totalStatistics.traditionalStatistics.pdo,
totalStatistics.traditionalStatistics.fullStrengthSavePercentagetotalStatistics.traditi
onalStatistics.fullStrengthShootingPercentage, matchid
5. AllMatchShots
o Source: Wisehockey API
o Contents: Detailed shot statistics for each shot in each match
o Features: team, period, result, shooter, expectedGoals, teamStrength.type,
matchid
6. AllPlayerStats
o Source: Wisehockey API
o Contents: Full season statistics for players
o Features: numberOfMatchesPlayed, playerid,
skatingStatistics.topSpeed,skatingStatistics.timeOnIce,
skatingStatistics.distanceTravelled,
skatingStatistics.averageSpeedWithPuck, skatingStatistics.accelerations,
skatingStatistics.decelerations, faceoffStatistics.faceoffWins,
faceoffStatistics.faceoffCount, plusMinusStatistics.plus,
plusMinusStatistics.minus, plusMinusStatistics.total, shotStatistics.shots,
shotStatistics.goals, passStatistics.successfulPasses, passStatistics.allPasses,
passStatistics.receivedPasses, passStatistics.totalPassDistance,
passStatistics.forwardPassDistance, puckControlStatistics.puckControlTime,
puckControlStatistics.puckControlDistance,
puckControlStatistics.puckControlForwardDistance,
puckContestStatistics.puckContestsWon,
puckContestStatistics.puckContestsLost,
puckContestStatistics.puckContestParticipations,

48
 shotScreenStatistics.offensiveScreens, shotScreenStatistics.blockedShots,
traditionalStatistics.corsiFor, traditionalStatistics.corsiAgainst,
traditionalStatistics.corsiForPercentage,
traditionalStatistics.relativeCorsiForPercentage,
traditionalStatistics.fenwickForPercentage,
traditionalStatistics.relativeFenwickForPercentage, traditionalStatistics.pdo,
traditionalStatistics.fullStrengthSavePercentage,
traditionalStatistics.fullStrengthShootingPercentage
7. SkatersLiiga
o Source: SportContract
o Contents: Full season statistics for skaters (all players excluding goalies)
o Features: id, firstName, lastName, position, teamId, teamName, jerseyNumber,
GP, G, A, P, PM, PlusMinus, PP, SH, AS, TOIPerGame, FOW, FOL, FO_perc,
BLS, SOG, Hits, A1, A2, PPG, SHG
8. SkatersHockeyAllSvenskan
o Source: SportContract
o Contents: Full season statistics for skaters (all players excluding goalies)
o Features: id, firstName, lastName, position, teamId, teamName, jerseyNumber,
GP, G, A, P, PM, PlusMinus, PP, SH, AS, TOIPerGame, FOW, FOL, FO_perc,
BLS, SOG, Hits, A1, A2, PPG, SHG

6.2 Analytics architecture

This chapter discusses the analytics architecture.

6.2.1 Architecture considerations

Key architecture considerations for analytics include:

• What data is needed to answer to the question?
• What data is available, where, and how to get it? How to ingest the data?
• Can the data answer to the business question? If not, is there more relevant data, or can
the question be modified?, Should exploratory analysis be used to find the insights from
that data?

49
 • What is needed to be done to the data before it can be used in analysis? How to transform
the data?
• Where to store the raw and transformed data?
• How to query and analyse the data?
• How to present the answer to the business question(s) in a format which is understandable
to human, so how to visualize and get meaningful insights and context to the data?

Organizations often have huge amounts of data, which is spread across different platforms and
services. As the amount of collected data is increasing, access to it becomes important. Many tools
offer in-application statistics and visualizations which may be too simple and are made only to fulfil
some simple task (Jarett, Kuo, 2021). Many companies have invested in data practices but still are
looking for how to make value from data. Becoming a data driven company requires a holistic data
strategy which among other things, considers the right ecosystem. Traditional way has been to
include point solutions which provide data services. However, due to the amount of the data, the
complexity, and the various needs and use cases these point solutions are not sufficient anymore
(Pierce, Tekiner, 2021). By centralizing data sources for cross-platform analytics enables
organizations to get an accurate picture of their business (Jarett, Kuo, 2021).

A platform with distributed data warehouse, which supports structured, semi-structured, and
streaming data, supporting clients and languages commonly used in data science such as Python,
R and SQL, including built-in capabilities for machine learning, and advanced analytics can be
called unified data analytics platform. (Paige, 2021)

Picture 1. Simplified architecture for data analytics.

50
Picture 1 illustrates a unified platform which is capable of ingesting diverse data from multiple
sources, has pipeline services, a data storage (the data environment), an analytics layer, and a
consumption layer. The platform is deployed in the cloud, is governed and secure by design.
(Halper, 2021)

6.2.2 Architecture and data processing flow

For this thesis purpose, I built a major part of the analysis tools and pipeline into AWS Cloud. I used
Amazon Web Services documentation at https://docs.aws.amazon.com/index.html as instructions
for building and configuring the environment and running analytics and “AWS serverless data
analytics pipeline reference architecture” for designing the architecture. Only for extracting the data
from both API sources, I used a Putin 3.8 with Jupyter notebook instance on my own laptop. Picture
2 presents the analysis pipeline and environment I built to ingest, store, transform, analyse, and
visualise the datasets. (Amazon Web Services Documentation)

Picture 2. Analysis pipeline in AWS.

The pipeline used for the thesis was:

• Data extraction from API - Microsoft Visual Studio Code with Jupyter Notebook plugin,
Python 3.8
• Data storage / Datalake - Amazon S3 bucket
• Raw data ingestion to datalake - Upload raw data files from my computer to S3
• Transforming raw data - Amazon Glue Jobs
• Importing data to data catalog for analysis - Amazon Glue Data Crawler
• Data catalog - Amazon Glue Data Catalog

51
 • Querying text data in datalake and building analysis tables and views - Amazon Athena
• Applying machine learning models - Amazon SageMaker Canvas
• Visualizing, extracting insights and detecting anomalies - Amazon QuickSight

Picture 3 below illustrates the flow for data processing used in this thesis. In this research,
steps 1 and 2 are done within researchers computer, steps 3 and 4 are done in the cloud.

Picture 3. Data processing flow.

1. Data is extracted from RESTful API interface with Python code and is stored in local
computer in JSON format.
2. Data is then combined and pre-processed into raw comma separated (CSV) datasets
presented in 6.1.5
3. Combined and pre-processed CSV datasets are uploaded into AWS S3 data lake (see
6.2.5 for details)
4. Datasets are transformed with AWS Glue ETL jobs into CSV and Parquet formats to be
queried and analysed and to be used in training and predictions with machine learning.

6.2.3 Extracting data from API

For this thesis I am using Jupyter notebook and Python to write custom code to fetch the datasets
from WiseHockey and SportContract API’s. As described earlier, both offer limited data via their
API, which indicates that to use advanced analytics already prepared by these sources, the data
needs to be ingested with different style; by using data scraping from their web portal.

For extracting and filtering the raw data for the thesis, I used Microsoft Visual Studio Code, Jupyter
notebook plugin and Python 3.8 to build the needed code on my laptop computer. Raw data
handling was done mainly using Pandas and Numpy libraries. Both APIs were using REST API

52
format and they had multiple query paths. Structure of API’s was such that no complex queries are
possible straight to the API meaning a larger portion of data needs to be extracted, and suitable
data filtered from the extracted raw datasets. The code is written so that it can be imported with
small modification to AWS Lambda functions for constant data fetching. Per season data needs to
however be fetched only once after the end of season. If dataset is needed to be updated and kept
current with most recent data, “season to date”, principle it can also be fetched regularly, for
example after each round of matches.

Data extraction was done in JSON (JavaScript Object Notation) and CSV (Comma Separated
Values) format. The Python code is portable and can be reused if such a pipeline is built for
production use later.

Other possibility is to used data scraping. It is a technique where scraping software basically reads
the same web interface, a web page, programmatically which is intended for the human eye. There
are multiple scraping tools available, both free open source and commercial licensed tools.
Scraping requires programming and the created scrapers can be affected by changes in structure,
naming and code of the target website. This means that they may require regular modifications and
updates. Another angle to consider with data scraping is license agreements and terms of
conditions. I did not investigate whether any of the two data sources allow or deny use of data
scrapers. (Wikipedia)

6.2.4 Ingesting data to datalake

Ingesting raw data into the data lake was done by simply uploading raw data files into raw folder in
a data lake. Further on, under the raw folder, raw data was organized to multiple sub-folders to
keep the data manageable.

Extracting data and ingesting it to the data lake can be done and automated in AWS by for example
using Lambda function written in Python. The Python code written for this thesis can be used as
template for the lambda function for future use. Lambda is serverless, meaning code can be
executed without running any servers. Lambda function can connect to the API interfaces over the
internet, query and extract the data from the API and upload the raw data to the data lake.

53
In AWS cloud-based data analysis solutions, this could be done by using for example Lambda
functions. Python code prepared for the thesis can be transferred to Lambdas.

6.2.5 Storing data

The data lake is a core component of the data analytics platform. Modern data warehouses support
self-service, advanced analytics, and data sharing. The data warehouse ingests multiple sources
of data and makes them available for analysis. (Halper, 2021)

For storing data, I used AWS S3 bucket as data lake which top level structure is shown in pictures
4 - 6. Data stored in S3 can be accessed with the services needed for transforming, analysing and
visualization. Raw data was uploaded to the data lake. The data lake used for the thesis has its
own folders for each dataset and use.

Picture 4. Data lake top level hierarchy.

Picture 5. Ingest folder in data lake containing raw uploaded datasets.

54
Picture 6. Transformed folder in data lake containing transformed datasets.

6.2.6 Processing and transforming

After ingestion, raw data is transformed and cleaned for processing. This is done with ETL tool
AWS Glue. Example ETL flow is shown in picture 7 below. As the data from both data sources are
pre-processed, it did not contain any major notable errors, thus can be used as such for the
analysis. In this phase, from raw data unneeded data features can be dropped, however this part
was done with the ingestion python code. Data was further transformed to CSV and Apache
Parquet format, which makes querying it more efficient and faster. In this phase datasets could also
be combined when needed. CSV format is mainly used for training Machine Learning models,
which require structured text format as input.

Picture 7. ETL flow in AWS Glue.

55
The results of the transformed dataset are then stored to the data lake and metadata is written to
AWS Glue data catalogue to enable querying it with Amazon Athena. Athena uses a SQL like
language to query the files in the data lake directly without setting any additional databases. With
Athena you can query the dataset with easily understandable and common language. The limitation
for Athena is that it is a query tool, and the underlying data files cannot be updated with it. It supports
using use Data Query Language commands (select), but not for example Data Manipulation
Language like update or insert. This made forming data tables and views more complex, because
I could not for example change values in underlying data or modify values. This suggests that for
a more complex environment, an SQL database, or Data Warehouse for may be needed for data
storage, update, and queries. Luckily, I was able to do most of the needed data modification tasks
within QuickSight, which again has SQL like language for queries and manipulation.

At the ETL phase the data can be cleaned and combined with other datasets. This is a
recommendation for future use cases, to use AWS Glue to combine and edit the datasets where
needed. However, because getting all the data and coding to extract and ingest the data, defining
structures, tables, and views, due to the structure of extracted data, I did not have extra time to
come back, code the ingestion and transformation again and I had to manage the scope by handling
the data in analysis and visualization phases.

6.2.7 Building queries and combining datasets

Data is queried and combined with Amazon Athena. I used Athena to build tables and views for
visualization. Table created table and view structure is presented in appendix 1.

At this stage complexity of exported and ingested dataset came visible. To calculate expected goal
statistic for a player in a single game, I needed to create some complex SQL queries which
combines multiple ingested datasets.

One of the issues is that not every player shoots or makes goals, and when this happens, that part
of the query per player is missing. One player could have played in more than one team in a season,
and in this case the SQL query join returns null value for the Players team. Amazon Athena is a
query tools and language, which means with it there is no possibility to update the underlying

56
datasets stored on S3 data lake. Previously described problems with the dataset could be avoided
by creating a SQL database and updating missing values with SQL query language. Using a SQL
database such as Amazon Aurora or Redshift data warehouse is recommended for future
architecture.

If the API would offer the data in the same structure which is displayed in the WiseHockey web-
portal, would also solve this, the table structure would contain only valid tables, the data ingested
and analysed data amount would be significantly less, and analysis could have been done without
any additional transformation and query building. At the current API exported dataset, we must
build all the advanced statistics, meaning we must build a portion of the logic of the application
again. Hockey teams already have access to those statistics through a web interface and are
paying a license fee for Wisehockey and can expect to have the advanced statistics accessible via
API.

6.2.8 Tables and views for analytics

For building the actual analytics selecting relevant features, combining, counting sums, and
querying the data is used following Athena tables. As the table structure is not relational, I am not
defining or drawing relation formulas nor drawing relation diagrams. Table 6 below describes
analysis tables and their content. Please see appendix 1 for detailed description of the tables. In
addition, I used a few temporary tables to build and test the final structure.

Table / view Contents Purpose

Liiga_2021_analysis_view_v1_ Final data for analysis Analysis and visualization
1 in QuickSight
2021allplayersandteams All players and their teams Interim phase table
2021allshots Data about all shots in all matches Query shot data
2021allteams Data about all teams Query team data
2021playerstats Data about players performance over the season Query seasonal
performance data
v_player_team_count Data about players team count during season Used for normalizing data
in visualization phase
2021allmatch Data about matches Query match data
2021allmatchplayerstat Data about players performance in a match Query match level
performance data
2021allplayers Data about players Query player data
Liiga2021allgamesgoals Combined data with goal information for all Query goal data
matches
Runkosarja2021allshotsxg Combined data with sum of expected goals for all Query expected goals
games for all players data
Rs2021playerswithteamsxg Combined data, expected goals data with players Interim data table for
team information building final view

57
 sc-liiga2021-skaters Performance data over the whole season from Train ML model
Sportcontract for Liiga
sc-allsvenskan2021-skaters Performance data over the whole season from Test ML model
Sportcontract for HockeyAllSvenskan
Table 6. Tables and views used for analysis.

6.2.9 Visualization

For visualizing and creating graphs from the data, I used Amazon QuickSight. QuickSight can use
the built Amazon Athena tables and views as a data source and it can be used to calculate
calculated fields, mean, max, median, and other statistics over the data. I used the built Athena
queries as datasets for the visualization.

6.2.10 Machine learning

For creating predictions on full season results, I used Amazon SageMaker canvas. SageMaker
canvas makes it possible to use machine learning with no coding. For this thesis purpose, to test if
I can predict season results, I used only quick model creation. The accuracy of the model increases
if model is further tuned, which means using methods requiring coding.

I trained the model with Liiga full 2021 regular season dataset and the trained model to predict
players performance in Liiga based on HockeyAllSvenskan 2021 full regular season dataset.
Because I only had a limited number of features and no advanced statistics at all in either dataset,
I trained the model to predict players scored goals in a season.

I used Canvas to try if time series prediction is possible based on season match data. This model
can predict players future performance in a defined timeframe from the prediction day onwards. As
I had only one full season of the data, the model can only give predictions 30 days into the future.
I trained a model to predict players' future goals, but the model can be trained to predict any relevant
feature which is present in the dataset. Example prediction is presented later in this thesis report.

For third use case for machine learning in this thesis, and with considering the results, I used
Amazon QuickSight Anomaly detection to detect unusual patterns in a player's performance. This
turned out to be an interesting approach and after discussion with the thesis target organization, is
approach they consider very interesting and see potential for future use. Instead of creating
predictions with complex statistics, we use simple features and data points from match data to

58
identify low and high performance. This model does not rely on any specific league, source, or
dataset. Any dataset from any league can be used to train a model to predict anomalies, meaning
if a player performs better, or worse, plays a longer period with making points or goals or without
goals, that can be detected as anomaly and automatic alert can be generated. This could give a
hint for scouts and coaches that a player is performing outside their normal performance,
regardless of the level, league, team, or role they are playing. Same analogy applies to time series
predictions, as I am not trying to compare players against others, between the teams or leagues,
and just trying to get an indication of the future performance. The relevant thing is that we have
enough data we can use to train the model and verify and make the predictions.

As advanced statistics may be important to monitor and to some extent predict players
performance, they only rely on data from history. The use of anomaly detection and machine
learning to identify anomalies and to predict future performance with simple match statistics is novel
and something to study more. In this model, I used to predict goals, because by making goals you
win games, but both cases could be used to predict any parameter in players in-game performance.

The initial trial to train a machine learning model to predict players performance in Liiga based on
their performance in other leagues is also something to study more in the future. With the model I
trained using this limited feature dataset is probably not very accurate (we cannot verify it
because we do not have data from future games in Liiga for all the players). With a more
extensive data feature set, models’ accuracy could be improved, however this still would give only
indication of the direction rather than accurate prediction. With more training data, the model gets
better and possibly would offer some value in player decisions. Data can be collected over time
on those players selected to play in Liiga and then used create a dataset based on their historical
data and real performance in Liiga. This can be then used to train a model. Collecting this data
could take time, and still the dataset would have too few data points to train accurate model.

Further on, an assumption can be made that an automatic machine learning models can indicate
what parameters are most relevant when they make predictions. This can give an indication what
features and statistics are most relevant to look after on that team, league, season, player, or
role.

59
Libraries like Autogluon can be used with SageMaker notebook instance, however, requires some
coding with python language. With Autogluon with only small coding effort a model can be
selected and trained and then perform predictions based on tabular dataset.

6.2.11 Security and data protection

Security and data protection of the data is important. The dataset used in this thesis contains
player names with their performance data, which is in big parts available on public internet, but
can be considered data which is regulated with GDPR. The analysis results also form a valuable
dataset. These two aspects are key considerations for security, and when such a system and
pipeline is put into production must be taken into use. For the thesis purposes, all data was held
in a data lake which is encrypted with strong encryption (using AES-256 with SSE-S3 keys) at
rest. Access to the data is limited to only to one analysis role and the service roles needed for
analysis and pipeline to run.

For the production analytics system, strong at rest and in transit encryption should be applied and
enforced, data lake policy to restrict access to the minimum, to enforce only encrypted uploads, to
restrict access from outside own organization and deny making data lake bucket public. It is
important to not open any access to the analytics environment, except Amazon QuickSight, from
the public internet. A fine-grained access policy can be created with AWS Lake Formation and to
ensure all components in the pipeline use encrypted data lake bucket as their data storage.

Further consideration is to use an additional layer of protection by protecting the data

environment with Threat Detection service like Amazon GuardDuty and identifying sensitive
information and possible misconfigurations with Amazon Macie.

To avoid unnecessary contractual and consent management it is important to store the data only
within EU. Further, the account used to store and process the data should be separate from other
purposes and locked down with proper access and service control policies, network connection
filtering and setting access boundaries.

60
6.3 Analytics results

6.3.1 Anomaly detection

When discussing with my thesis principal KalPa, one of their specific wishes was to find out if
there is a way to automatically detect unusual patterns in players performance, in other words is
there a way to detect anomalies and get alerted automatically when they are detected.

Anomaly detection could give indication on changes on players performance, for example if a
player is on longer strike making points. I used Amazon QuickSight for visualizing the results, and
that service has built in automatic capability to use Machine Learning for detecting anomalies in
the data set. (Amazon)

Anomaly detection has been well studied over the last few decades and is one of the key
problems in data mining. Key questions in detecting anomalies are 1) how anomaly is defined
and 2) what data structure support to detect anomalies? Generally, it can be said that a data point
is an anomaly if the complexity of the model increases substantially with the inclusion of the data
point. In supervised learning randomization has been found to be effective l tool. Guha, Misra,
Roy and Schrijvers introduced the robust random cut forest sketch. The Robust Random Cut
Forest (RRCF) algorithm is an efficient ensemble method for detecting outliers in streaming or
time series data. (Guha, Mishra, Roy, Schrijvers, 2016)

A random cut forest (RCF) is a special type of random forest (RF) algorithm which takes a set of
random data points, cuts them down to the same number of points, and then builds a collection of
models. RCF is similar to a decision tree, is an unsupervised algorithm and uses cluster analysis
to detect spikes in time series data, breaks in periodicity or seasonality, and data point
exceptions. (Amazon)

Amazon QuickSight uses RCF for anomaly detection. In addition to detection, it includes a feature
to send automatic alerts when anomaly has been found (Amazon). As a test case for anomaly
detection, I used a ranking value I calculated using expected goals and PDO. Picture 8 shows
anomalies QuickSight identified on data for xG Rank March 2022 and picture 9 detailed view of
first identified anomaly.

61
When looking at the absolute values identified in the anomalies, it should be noted that the Rank
xG is reverse rank, meaning the lower value is better. Here It can be observed that for example
Niclas Lundgred played better than usual, hence an anomaly was detected.

Picture 8. Anomalies on Rank xG on March 22, 2022.

When looking closer at the anomaly on Lundgren, the graphical interpretation of the anomaly can
be used. The graphic is scaled, and value should be multiplied with 1000. From the graphic it is
observed that Lundgrens xG rank per game have varied around 20000 with regular spikes closer
to 1000, which all would have been detected as anomalies. On March 22nd the value was again
detected as anomaly, meaning Niclas Lundgrens expected goals in a game was higher than
expected, which could indicate better than normal performance.

Picture 9. Detailed view of anomalies for Niclas Lundgren.

As another sample for anomaly detection, I used PDO (picture 10). On Jan 29, 2022, an anomaly
was detected for Juuso Könönen and Reid Gardiner. PDO for Könönen was lower than usual,
and for Gardiner higher than usual. For Könönen this could indicate lower performance than
expected, and for Gardiner it can indicate higher performance than expected.

62
Picture 10. PDO anomalies for Könönen and Gardiner.

What is notable here is that these anomalies would have been automatically detected and
coaches and scouts alerted.

6.3.2 Normalising and summarising

In the dataset extracted from Wisehockey, when a player has been playing in more than one
team during the season, a player has a player entry for each team. For example, Reid Gardiner
played in two teams during regular season 2021-2022 (JYP and HIFK). This fact caused that in
the analytics tables for shot based statistics I created using Amazon Athena queries, such players
had multiple entries for each match. This problem was overcome by normalizing the summary
values with the count of players teams in that season. This method was used to normalize
summarized expected goals and goals values.

6.3.3 Expected goals ranking

For evaluating players performance, I created a statistical value, Rank xG, which gives a ranking
for players performance on a particular game using expected goals. So, for each player in each
game where the player made a shot for which expected goal value was given a ranking value
was given. This value is descendant, meaning value 1 is the best, 2 second best and so on. By
doing so, expected goal statistics in a time series is obtained. This means that players'
performance against others using time as a variable can be valuated and that the ranking number
directly indicates the order. The highest on the list is the best of all records during regular season
2021-2022. An absolute expected goal value on a given day can be used, but that would not
indicate its relation to other players in a single value. This metric gives a single number view to
players performance; it measures the offensive pressure a player has when on ice compared to

63
the reference group. As a reference group, the whole Liiga, players own team or players role or a
combination of the earlier can be used. Table 7 shows a sample of calculated rankings by using
the whole Liiga as a reference group.

Table 7. xG Ranking.

I compared three players' performance using the calculated Rank xG (Reid Gardiner, HIFK,
Juuso Könönen, KalPa and Aapeli Räsänen, KalPa). In addition, I used Rank xG ranking in
anomaly detection. When looking at the results on picture 11 should be noted values in y axis
should be read as absolute values. Picture shows that a team's coach can possibly use this data
as one data source, but for many without detailed understanding of the game on a particular date,
it is hard to identify the patterns for better or lower than usual performance. Hence using ML
based anomaly detection makes sense.

Picture 11. Rank xG for Gardiner, Könönen and Räsänen.

64
6.3.4 Analysis against reference group

As defined in earlier of this study, players' performance can be compared against each other,
when compared within their own reference group. In this example analysis, I used a comparison
between three Liiga forward skaters, Reid Gardiner (HIFK), Juuso Könönen (KalPa) and Aapeli
Räsänen (KalPa). Comparison visualizes the development and performance over the comparison
period, which in this example is Liiga full regular season 2021. From the visualization it is easy to
see stages during the season, how players perform during the season and the comparison over
the full season. However, it should be noted that playing in different level teams, where better
teams win more often, thus players are likely to score more often, should be considered in
comparison. But for example, comparing two players within the same team should give reliable
indication of their differences.

Picture 12 presents goals players have scored and picture 13 expected goals for a player during
the comparison period. It is easy to see the performance difference between players. However,
when looking at expected goals comparison, observation is made that both KalPa players have
played on the same level almost for the whole season.

Picture 12. Shots resulting a goal.

Picture 13. Expected goals.

65
Visualization of plus/minus points for the same players is a little bit different. As discussed earlier,
plus/minus point is given for a player if his team makes goals, or if the opponent team makes
goals when a player is on ice (Kohl, 2016). In this comparison observation is made that even if
Gardiner has scored a large amount more goals than Räsänen, in the second half of the season
their plus/minus statistics are on the same level (picture 14.) This indicates that it is just how
many goals a player scores, but how he contributes to the whole team's performance, so what
happens when the player is on ice. It is visible from the visual that the player's plus/minus statistic
has been decreasing in the first half of the season and been more stable during the second half.

Picture 14. Plus/minus.

In this dataset plus/minus is a normally distributed having centre in 0 (picture 15.) This means
that for each game data, most of the players play with zero plus/minus statistics, which gives an
indication that performance is better or worse with players who are outside the centre. In a single
game, higher plus/minus value can indicate better performance and negative value lower.
Interestingly, when a sum is calculated over the season, the better players seem to have a
negative plus/minus sum value. This can indicate that they also play more short-handed shifts,
and that they simply are more on ice, so the likelihood of the opponent goals during their shifts is
higher.

66
Picture 15. Distribution of plus/minus.

Picture 16 shows visualization of the same three players and their PDO during a season. As
defined before, average players PDO should be around 1 (or 100 depending on the scale), above
that is better performance and lower values indicate lower performance. This means that PDO is
a statistic that is not feasible to use in sum calculation. It is not easy to see the differences over
time, PDO is one of the statistics which would benefit from using machine learning based
anomaly detection in detecting unusual performance patterns. Picture 16 has been zoomed to
illustrate the second half of the season. When looking at how PDO values are distributed in the
dataset, it can be observed that it is close to normal distribution, having centre in value 1 (picture
17.) This is aligned with the theory part. This also can indicate that better player PDO is more
than 1 and lower performance players value is below 1.

Picture 16. PDO.

67
Picture 17. PDO distribution.

When visualizing the expected goals value (xG) over the season (picture 18), similar effect with
PDO is seen. It is harder to spot how players performance changes during season, and this time
series statistic could be used with anomaly detection. There are spikes in the data for players,
which likely will be identified as anomalies. Anomaly detection can identify better or lower
performance periods for a player. Using this with the sum over season statistics, and xG Rank
discussed earlier can indicate players performance development direction. The distribution for xG
in picture 19 is close to exponential cumulative distribution function (CDF), meaning most of the
values in the dataset are close to zero, and probability of higher values follow the CDF, the
probability approaches zero when the value of the x gets higher. Anomalies may be seen values
higher or lower of the function value. From this analysis, it is also possible to see streaks of better
and lower performance.

Picture 18. xG by game date.

68
Picture 19. Distribution of xG.

Another way to get indication of performance is to compare a player against the reference group,
for example own team. Picture 20 visualizes how Könönen and Räsänen xG values per game are
situated in comparison to Liiga average and own team (KalPa) average on a given game day.
Picture 21 takes into consideration players per role, it compares the players against forwards.

Picture 20. Expected goals comparison.

Picture 21. Expected goals comparison, forwards only.

69
Picture 22 does the same for PDO. In this analysis, it is observed that as PDO average is close to
one there is not much value for calculating the average to different groups for comparison. From
this analysis, some game dates are highlighted where these two players have been more lucky or
unlucky.

Picture 22. PDO comparison.

Corsi For Percentage (CF%) is the sum of shots on goal, missed shots, and blocked shots over
the shots against, missed shots against and blocked shots against at equal strength evaluates a
player’s or teams puck possession. A CF% value of 45-55% is typical to a hockey player
(Puckpedia). In other words, CF% values over 55% indicate better performance, and values
below 45% lower performance. Picture 23 indicates CF% comparison for Könönen and Räsänen,
as well as comparison against all players in Liiga, and all players in their own team. CF% value
for all players in Liiga seems to be close to 50% during the whole regular season 2021-2022. We
can see that CF% for Könönen and Räsänen varies above and below leagues and own teams
average over the season.

Picture 23. Corsi For%.

Yet another way to illustrate performance is to draw a scatter plot. Machine learning could be
applied to find patterns and grouping. Picture 24 and 25 presents a sum of expected goals value
in relation to time on ice. Picture 24 presents only players in defence roles and picture 25 the
players in forward roles. Higher performing players are in the top right.

70
Picture 24. xG and TOI for defence.

Picture 25. xG and TOI for forwards.

I trained a machine learning model with Amazon SageMaker Canvas autoML to make a time
series prediction on Goals. I used an exported dataset from the data I had ingested in the data
lake and let SageMaker decide which parameters are important to make such predictions. No
further fine tuning for the model was made.

71
The experiment shows that I was able to make a prediction for 30 days into the future. As a test
case, I used two players, Aapeli Räsänen (KalPa) (picture 26) and Reid Gardiner (HIFK) (picture
27.) KalPa did not play in playoffs during season 2022 so Räsänen prediction cannot be
compared to the real goal amount. When looking at the prediction for Reid Gardiner, HIFK, and
comparing the result to the 2022 playoffs the model predicted 1 goal for Gardiner, and he scored
2 goals during the playoffs (source Liiga.fi). As the data amount used in my thesis is limited to
only regular season 2021-2022, and does not include data from previous seasons playoffs, it is
not meaningful to continue to interpret the results with other players. The key takeaway is that
training a model to create time series predictions is possible, and with tuning the model and with
correct data, the model could give indicative predictions on players future performance.
Furthermore, a model can be trained to create predictions for the regular season as well. To keep
the scope of my thesis in control, I did not dive deep into this topic, and this alone would be an
idea for another master’s thesis, or even doctoral dissertation.

Picture 26. 30-day prediction for Aapeli Räsänen

Picture 27. 30-day prediction for Reid Gardiner

I trained SageMaker AutoML to predict performance for HockeyAllSvenskan players. I used full
season statistics from SportContract API for Liiga from regular season 2021-2022 to train the
model, and regular season 2021-2022 data for HockeyAllSvenskan players to create predictions.

72
I used this dataset because the features for both datasets are equal, and one can be used to train
a model and the other as testing the model. With the predictions we see that the model predicted
7.3 goals for Colby Sissons if he is playing in Liiga. In the regular season 2022-2023 Sissons had
made 8 goals by Feb 5th, 2023. It should be noted that predictions were made with earlier season
data and are not directly comparable to the current season. Sissons performance in Liiga can be
on a higher level so cannot directly be compared. However, the prediction for this one case could
be seen as an indicative prediction.

This test indicates following of findings:

• Dataset is too small to train the model for reliable prediction. Dataset should contain
multiple seasons to have enough data points.
• With having only one reference point which can be used to check the accuracy against
real results, it is impossible to evaluate the reliability. However, this could give indication
how a player would perform in Liiga.
• The most important finding is that results directly from different leagues should and
cannot be directly compared due to differences in playing styles, levels, and players, but
patterns from the data can be identified in any of the leagues where similar data is
available. These patterns, like anomalies, are not directly related to absolute data values
and possibly can be used for comparing players from any league. By using machine
learning in a similar way as I did in this experiment, would probably not give accurate
prediction on real performance, but can give indication on the direction of a level of a
player.

73
Table 8. Predicting players' scored goals.

Picture 28. Trained model statistics.

74
7 CONCLUSION AND FINDINGS

The aim and main objectives are to create both data and statistical models for analysing a hockey
game data in relation to the defined indicators for identifying better than normal player
performance. In addition to the creation of the models, the objective was to run the analysis with
real life data.

Secondary objective was to identify relevant indicators which can be used to identify better than
normal and exceptional performance of hockey players in their career by analysing the available
game data.

I first did qualitative literature-based research to find commonly known metrics and statistics used
in analysing hockey players performance. In the second part developed a data and statistical
model for analysis, built an analysis pipeline and finally used real life data for Liiga and
HockeyAllSvenskan to create example analysis. Data was exported from Wisehockey and
SportContract API’s. Analysis pipeline was built using Amazon Web Services data analytics
capabilities. In addition, I studied the use of machine learning for the evaluation and predictions.

7.1 Findings

First research question was to find out how to use data and analytics to identify players whose
performance development is better than normal and to identify the indicators, statistical model
and data which can identify better than usual and exceptional players (outliers).

My research shows that by analysing game data from the matches teams can get insights about
both teams and players performance. This thesis focused on individual players' performance.
Traditional statistical analysis methods can be applied to get insights and information to support
teams’ development and decision making. By applying analytics, a team’s management and
coaches can find optimal combinations for their player roster. This can be optimized from a player
budget perspective.

In addition to traditional statistics, machine learning can be applied to get insights. In this thesis, I
build an example of anomaly detection using Random Forest Cut. With anomaly detection teams

75
can identify exceptional performance and outliers (anomalies) in players performance which could
indicate better or worse than normal development and periods.

I then used time series data to build predictions for players' performance into the future. This
could be an interesting approach, to predict how a player or a team will perform in the series of
future games, such as in playoffs. This however requires enough data and context to give
trustworthy predictions. With only one season worth of data predictions will not be accurate
enough. This however could be a future research topic.

I studied using a machine learning model to see if I can train a model with target leagues data
and predict another leagues players performance with it. Because the game and playing styles
differ between leagues, it cannot give accurate predictions, however, can give indication how a
player from another league would compare to another leagues level. Also, this topic be an
interesting for future research.

Second research question was to find out what advanced metrics and statistics models are
common in hockey, and will they fit for analysing Liiga and Allsvenskan game data and what
additional data would be needed to use these metrics and models?

During my research I found out that there are numerous methods and advanced statistics
commonly known which all try to evaluate and to predict performance or value. Many of these are
targeted to evaluate performance on a team level and on individual player level. Some are
complex and their value in analysis for Liiga and HockeyAllSvenskan data could not be proven
within the scope of this thesis. Many of them require special types of data, such as NHL RTSS
data, and are targeted to evaluate performance in the NHL and most of them were not available
in the data sources accessible to me during my research. There are statistics metrics that can
used to evaluate and predict performance in Liiga and HockeyAllSvenskan. These statistics
include, for example, goals, points, shots, PDO, faceoffs won/lost, time on ice, plus/minus,
assists, expected goals, Corsi and Fenwick.

I used a model which combines goals, expected goals, plus/minus, time on ice, Corsi and PDO.
In addition to these I calculated an expected goals rank over the full season. This metrics gives a
ranking for all the players in all the games during a season using xG sum of their shots in a game.
This metric gives a single number view to players performance; it measures the offensive

76
pressure a player has when on ice compared to the reference group. As a reference group, whole
Liiga, players own team or players role or a combination of the earlier can be used.

Third research question was to find out what data is available to support analysis and decision
making which could be used as a source for automated data pipeline and continuous analysis?

My research shows that there are both publicly available and free web sites and commercial and
non-public API sources where the data can be obtained. I used two APIs to fetch the data for
analysis, Wisehockey and SportContract. Both are RESTful APIs and relatively easy to use
programmatically with for example using Python and Python libraries and both offer simple
documentation of the API contents and structures. Data from public web sites could be extracted
by for example using data scraping.

7.2 Hypothesis testing

The hypothesis for the research is that by using commonly known ice hockey statistics to analyse
game data and by using machine learning models, a professional hockey team's management,
coaches and scouts can identify and predict players performance and use it to support their
player recruitment decisions.

My research has indicated that data analytics is common in professional sports and becoming
even more important and relevant in ice hockey. There are data which can be used for analysis,
tools and methods are commonly known and available. In addition, this research shows that
machine learning can help to identify unusual (better or worse than normal) player performance
and may support to predict future performance. If a player's performance or potential is evaluated
without analysing the supporting data, the view is always biased and subjective. This research
shows that data analysis can play a key part in identifying players' performance and provides
examples how the data can be used to predict their performance in the future. In leagues like the
NHL, using data in the player recruitment process is common and multiple methods and statistics
are available to support the decision making. Thus, the hypothesis is true.

77
7.3 Other findings

An additional finding of this research is related to comparison of players playing in different

leagues. As there may be significant differences in the playing style, skill level, and rules, the data
and performance between two leagues are not directly comparable, which means the data from
the two can't be directly compared. When you find patterns from the data with for example
machine learning, like anomalies or other metrics or performance identifiers (like the xG rank
used in this thesis), this new data is not anymore directly related to the original data or the
leagues where it originates. If generated by identical means, this new data is then naturally
normalized and becomes comparable.

Openness of the data is also a finding worth mentioning. In NHL the data seems to be more
easily and openly available to anyone who wishes to do analytics, for both professionals and
hobbyists. If the data platforms would be more open it would allow data owners, data consumers
and third parties to interact and contribute their feedback. (de Reuver, Ofe, Agahari, Abbas,
Antragama, Zuiderwijk, 2022)

If data in these researched leagues would be more openly available, it could create both new
business opportunities, make the data platforms better with help of the community and create
opportunities for players and teams to utilize the data in more efficient ways.

The analysis case in thesis was a good learning experience and it covered the full lifecycle of
analysis. It started with data engineering, had phase for data analysis and data science and
finally ending into visualizing business insights and presenting them to the stakeholders.

78
8 DISCUSSION

8.1 Challenges, issues and resolutions

When evaluating the results against the data, I noticed that there are some inconsistencies with
the datasets. This can be seen for example in the analysis of goals, the results show that Juuso
Könönen has made 11 goals in the analysed time period. However, Liiga.fi statistics indicate that
the actual number was 12. When I checked this from the exported raw dataset, it has 11 shots for
Könönen which has resulted in a goal. This could mean that for example the problem appears
during data extraction from Wisehockey API or error in the data source which then results in a
completely missing data entry. One possibility for the error is that I have calculated the goals
using shot data, any shot with the result of goal is counted as goals. But if there are goals
recorded which did not originate from a shot, those are then not calculated. I checked this from
Wisehockey API documentation, and the goal statistic references the shot, so this should be a
correct way to calculate. I did not investigate this further, but it should be fixed before doing any
real analysis with the data.

An issue I found with the summarized datasets was that because a player could have been
playing in a match where they did not shoot any shots, and as result had missing values in goals
and expected goals values. This has an effect when the average over a group of players is
calculated and compared that to a player. This error in the dataset influences the averages, but
as such does not prevent nor introduce significant bias on finding answers to the research
questions. This error however in the data model should be fixed if a similar model would be taken
into real analysis use. Easy fix could be to design and import the data into a relational database,
where the data can be then easily queried and missing values altered. Because this error did not
actually affect finding answers to the research questions and considering that data handling took
much more effort and time than planned, I decided not to design relations and not to import the
data from S3 based data lake to a database. Amazon Athena, which I used for building the
analysis tables and views, is a flexible and easy tool for analysing and querying data on
structured files. Its limitation however is that it does not allow to be able to directly alter the
underlying data files.

79
Another solution could be to combine datasets in a transformation phase. This however would
have been a step back in the process and would have required again too much work to handle
within the scope of this thesis. Initially, I did not do this in the transformation phase because I had
to explore the data and find out if I could answer the research questions with it.

Yet another challenge I faced is related to the data itself and getting the data from the API’s. As
mentioned earlier the two sources used both had RESTful API to access the data
programmatically. SportContract API contents were easy to extract yet had very limited datasets.
Wisehockey API had more data available, however advanced stats they offer via web interface
were not among the data. API from Wisehockey was more complex and I needed to extract and
handle, transform and query much more data which I in later phases needed to summarize and
combine to get my hands into the wanted statistics, like goals and expected goals. The data was
rather raw data. As I am not a professional programmer, I had a steep learning curve in coding to
extract the data. The fact that I needed to do more coding work in exporting the data meant I
needed to combine and summarize and calculate the features in the analysis phase. This turned
out to be much more laborious than I expected in the thesis planning phase. I estimate that these
two elements together introduced 3 - 4 weeks of more work than what was planned. To keep the
thesis scope manageable, I needed to narrow down the statistics I used in the analysis and to do
the machine learning part more superficially by only testing concepts and building some
examples. However, this does not have a significant effect on the thesis results when assessing
them against the research questions and goals. What it means is that I did not get to the analytics
automation part as deeply as I wanted to.

I assumed the data sources are providing already cleaned and accurate data which can be
ingested into analytics without too much preparation, cleaning, and transformation, and that the
data sets are clean and complete. This was not the case for all parts. This challenge was
highlighted by the fact that I needed to summarize and calculate statistics from raw data and
structures of the exported data. This leads to more work and questions with for example missing
value handling, duplicate data entries showing after combining and summarising. This problem
was discussed earlier in the thesis and the easiest solution to avoid such difficulties is to use a
database where data can be altered as needed. This caused some confusion and extra work but
does not provide a significant negative impact on thesis results. Even if the ingested data looks
complete, the need for cleaning, supplementing and imputing values to the dataset in the analysis
phase should be considered.

80
Other things to consider with such comprehensive thesis topic and with doing thesis part time
along with full time day job, is the thought work it requires. Catching up from the previous day’s
work, doing a small piece of work, saving all work done and then wrapping up to start same
process again the next night is not very effective way. This introduced some limitations on the
analysis scope.

8.2 Data and results reliability

Neither of the sources publish publicly the information about the error margins and reliability of
the raw data, metrics, and statistics nor how specifically the specific metric is collected or
calculated. There has been debate in the public news about the data trustworthiness and
inaccuracies in the metrics and the data in Wisehockey platform (MTV Uutiset). This news article
was published 16.10.2022 and I exported the data from the API 14.11.2022. My research
indicates some inconsistencies which could be caused by the data source but also by a process
used for data extraction. In addition, both used data sources limit the statistics available through
their API, which means a solution must be developed to calculate these statistics with raw data if
those will be used in the analysis. This introduces error possibility which could have a significant
effect on the results and in the worst-case lead to wrong decisions.

This means that the good as the analytics capabilities in the platform are, the results should be
seen as indicative until source data correctness and trustworthiness can be confirmed. This also
means that teams should continue to use subject matter expertise to back up their decision
making.

8.3 Future research possibilities

During my research I identified several additional questions to be researched which I could not
answer, or they were not in scope for this thesis. These could be for example studied in other
thesis projects or used as an idea for doctoral studies and dissertation. Potential questions and
issues are:

1. Would hockey leagues like Liiga benefit from more open data strategy and would it
provide opportunities for teams and players?

81
2. How to apply novel methods such as deep learning, neural networks, and computer
vision in hockey analytics?
3. Developing and using machine learning models to identify performance in near real time
during a match and using video stream analytics.
4. More in-depth research on using machine learning and anomaly detection-based
predictions.
5. Using machine learning to evaluate and predict performance between different leagues.
6. Using machine learning to predict season players and teams’ performance in ongoing
season and use the model to build optimal team compositions.

82
9 ACKNOWLEDGEMENTS

I would like to acknowledge and give my warmest thanks to my thesis supervisor Dr. Teppo
Räisänen. With his guidance I was able to successfully go through all the stages of writing my
thesis. I would also like to give special thanks to Anssi Laine and Atte Sacklin and the whole
KalPa Hockey team for providing this opportunity to research this very interesting topic.

Special thanks to Dalton Linkus and Janne Kekäläinen from Nashville Predators for their views
and insights and to Sami Knuutinen for tipping me this opportunity.

83
REFERENCES

Amazon Web Services Documentation. https://docs.aws.amazon.com/index.html

Brower, Owen 2020. Applying Analytics to the NFL Draft: Athletic Performance Measures as a
Predictor of Future Success. Economics Theses. Search date: 6.9.2022.
https://soundideas.pugetsound.edu/economics_theses/111

bgred105, 2012. Sab(re)metrics: What is GVT? Search date: 16.10.2022.

https://www.diebytheblade.com/2012/12/7/3411686/sab-re-metrics-what-is-gvt

Bitwise Oy. Wisechockey. Search date: 22.2.2023. https://bitwise.fi/

Burtch, Steve. 2014. What is dCorsi and how does it help us account for "context”? Search Date:
12.10.2022. https://www.pensionplanpuppets.com/2014/7/21/5663440/dcorsi-introduction-what-
is-dcorsi-how-do-you-use-dcorsi

Burtch, Steve. 2014. Delta Corsi and Assessment of Individual Player Impacts on Possession
Accounting for Usage. https://www.dropbox.com/s/wb10dq5mlpv0uzz/dCorsi-WRITEUP.pdf

Cambridge Dictionary. Search date: 9.9.2022.

https://dictionary.cambridge.org/dictionary/english/data-analysis

de Reuver, Mark & Ofe, Hosea Ayaba & Agahari, Wireman & Abbas, Antragama Ewa & Zuiderwijk,
Anneke. 2022. The Openness of Data Platforms: A Research Agenda. Search date: 2.3.2023.
https://www.researchgate.net/publication/365122960_The_Openness_of_Data_Platforms_A_Res
earch_Agenda

Gautam, Akshay 2021. 6 Phases Of Data Analysis According To Google. Search date: 9.9.2022.
https://medium.com/codex/6-phases-of-data-analysis-according-to-google-9e084b89f848

CLVNNG, 2017. Calculating Hockey Analytic’s GVT. Search date: 12.10.2022.

https://performancedrivenanalytics.wordpress.com/2017/01/04/34/

84
Found, Rob. 2016. Goal-based Metrics Better Than Shot-based Metrics at Predicting Hockey
Success. Search date: 17.12.2022. https://www.researchgate.net/publication/325302195_Goal-
based_Metrics_Better_Than_Shot-based_Metrics_at_Predicting_Hockey_Success US Sports
Academy.

Halper, Fern, 2021. Unified Platforms for Modern Analytics. TDWI.

https://tdwi.org/research/2021/09/arch-all-best-practices-report-unified-platforms-for-modern-
analytics.aspx?tc=page0

Jahn K. Hakes & Raymond D. Sauer 2006. An Economic Evaluation of the Moneyball Hypothesis,
Journal of Economic Perspectives, American Economic Association, vol. 20(3), pages 173-186,
Summer. Search date: 8.9.2022- https://ideas.repec.org/a/aea/jecper/v20y2006i3p173-186.html

Jarett, Leigha & Kuo Micheal, 2021. Creating a unified analytics platform for digital natives. Google.
Search date 5.2.2023. https://cloud.google.com/blog/topics/developers-practitioners/creating-
unified-analytics-platform-digital-natives

Jones, Wayne. What does plus minus (+/-) mean in hockey? (with stats). Search date: 10.10.2022.
https://hockeyanswered.com/what-does-plus-minus-mean-in-hockey-with-stats/

Joshua Weissbock & Herna Viktor & Diana Inkpen. Use of Performance Metrics to Forecast
Success in the National Hockey League. University of Ottawa, Ottawa, Canada.

Julie, Kathleen 2014. The Many Ways the NHL Is Using Data Collection This Season to Enhance
the League. Search date: 2.9.2022. https://www.sporttechie.com/the-many-ways-the-nhl-is-using-
data-collection-this-season-to-enhance-the-league

Junior KalPa web page. Search date: 2.9.2022. https://kalpa.fi/en-gb/article/etusivu/junior-

kalpa/43/

KalPa Hockey web page. Search date: 2.9.2022. https://kalpa.fi/?lang=en-gb&langmenu=1

85
Kohl, Garret, 2016. Behind the Numbers: Why Plus/Minus is the worst statistic in hockey and should
be abolished. Search date: 5.10.2022. https://hockey-graphs.com/2016/11/01/behind-the-
numbers-why-plusminus-is-the-worst-statistic-in-hockey-and-should-be-abolished/

Kohonen, Iina & Kuula-Luumi, Arja & Spoof, Sanna-Kaisa 2019. The ethical principles of research
with human participants and ethical review in the human sciences in Finland Finnish National Board
on Research Integrity TENK guidelines. Search date: 15.8.2022.
https://tenk.fi/sites/default/files/2021-01/Ethical_review_in_human_sciences_2020.pdf

Kontsas, Jukka & Lehtola, Juha 2014. Goalie and scoring analysis: MOL, Mestis and Liiga. Haaga-
Helia University of Applied Sciences. Search date: 5.9.2022.
https://www.theseus.fi/bitstream/handle/10024/78338/Kontsas%20Lehtola%20Final.pdf?sequenc
e=1

Kubatko, Justin. 2011. Calculating Point Shares. Search date: 11.10.2022. https://www.hockey-
reference.com/about/point_shares.html

Larkin P, O’Connor D 2017. Talent identification and recruitment in youth soccer: Recruiter’s
perceptions of the key attributes for player recruitment. PLoS ONE. Search date: 5.9.2022
https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0175716&type=printable

Lee, Chirstian, 2020. Advanced Hockey Stats 101: Corsi. Search date: 11.10.2022.
https://medium.com/hockey-stats/advanced-hockey-stats-101-corsi-part-1-of-4-29d0a9fb1f95

Lee, Chirstian, 2020. Advanced Hockey Stats 101: Fenwick. Search date: 11.10.2022.
https://medium.com/hockey-stats/advanced-hockey-stats-101-fenwick-part-2-of-4-a8796b8188d

Lee, Chirstian, 2021. Advanced Hockey Stats 101: PDO. Search date: 11.10.2022.
https://medium.com/hockey-stats/advanced-hockey-stats-101-pdo-part-3-of-4-d3f319f2e1f1

Lee, Chirstian, 2021. Advanced Hockey Stats 101: Zone Starts. Search date: 11.10.2022.
https://medium.com/hockey-stats/advanced-hockey-stats-101-zone-starts-part-4-of-4-
1548b894541c

86
Liu, Guiliang & Schulte, Oliver 2018. Deep Reinforcement Learning in Ice Hockey for Context-
Aware Player Evaluation. Simon Fraser University, Burnaby, Canada. Search date: 7.9.2022.
https://arxiv.org/pdf/1805.11088.pdf

Lukas, Allison. 2021. Analytics with Alison: Expected Goals. Search date: 16.11.2022
https://www.nhl.com/kraken/news/analytics-with-alison-expected-goals/c-327728890 Seattle
Kraken.

Macdonald, Brian 2011. A Regression-Based Adjusted Plus-Minus Statistic for NHL Players,
Journal of Quantitative Analysis in Sports: Vol. 7: Iss. 3, Article 4. Search date:
9.9.2022. http://hockeyanalytics.com/Research_files/Regression_Based_Plus_Minus.pdf

MacDonald, Brian. 2012. An Expected Goals Model for Evaluating NHL Teams and Players.
http://www.hockeyanalytics.com/Research_files/NHL-Expected-Goals-Brian-Macdonald.pdf MIT
Sloan Sports Analytics Conference 2012.

Mondello, Michael & Kamke, Christopher 2014. The Introduction and Application of Sports
Analytics in Professional Sport Organizations. Journal of Applied Sport Management: Vol. 6 : Iss.
2. Search date: 3.9.2022.
https://trace.tennessee.edu/cgi/viewcontent.cgi?article=1357&context=jasm

MTV Uutiset. 2021. SM-liigan tilastointi huippuluotsien hampaissa – pelkästään tämä esimerkki
kertoo raa'asta epätarkkuudesta: "Aikamoisia numeroita”. Search date: 12.2.2023.
https://www.mtvuutiset.fi/artikkeli/sm-liigan-tilastointi-huippuluotsien-hampaissa-pelkastaan-tama-
esimerkki-kertoo-raa-asta-epatarkkuudesta-aikamoisia-numeroita/8549030

Paige, Robert, 2021. What’s the Difference Between a Data Lakehouse and a Unified Analyt-ics
Platform? Architecture & Governance Magazine. Search date 5.2.2023.
https://www.architectureandgovernance.com/digital-transformation/whats-the-difference-between-
a-data-lakehouse-and-a-unified-analytics-platform/

Pallotta, Jack 2021. Goals for per 60: The Best Of NHL Statistics. Search date: 14.9.2022.

87
Piccolo, Nick 2022. Inside the Stats: Corsi, Fenwick and On-Ice Stats. Search date: 14.9.2022.
https://insidetherink.com/inside-the-stats-corsi-fenwick-and-on-ice-stats/

Pierce, Susan & Tekiner, Firat, 2021. Building a unified analytics data platform on Google Cloud,.
Search date 5.2.2023. Google. https://cloud.google.com/blog/products/data-analytics/building-
unified-analytics-data-platform-google-cloud

Pettigrew, Stephen 2015. Assessing the offensive productivity of NHL players using in-game win
probabilities. Harvard University. Search date: 6.9.2022. https://global-
uploads.webflow.com/5f1af76ed86d6771ad48324b/5fee098cacad0134c9e8d403_1028_rppaper_
pettigrew%20sloan%20submission.pdf

Praful Kava and Changbin Gong. 2020. AWS serverless data analytics pipeline reference
architecture. Search date: 1.11.2022. https://aws.amazon.com/blogs/big-data/aws-serverless-
data-analytics-pipeline-reference-architecture/ Amazon.

Puckpedia. NHL Player Stats. Search datę: 12.1.2023.

https://puckpedia.com/stats#:~:text=Corsi%20For%20Percentage%20(CF%25),shots%20against
%20at%20equal%20strength

Raber, Matthew & Eisenberg, Daniel. Defensive Efficiency Metrics (DEMs): A Paradigm Shift in
Defensive Hockey Analysis Based on Per-Possession Measurements. Big Data Cup.

Ryder, Alan. 2008. Ten Laws of Hockey Analytics. Search date: 10.10.2022.
http://hockeyanalytics.com/2008/01/the-ten-laws-of-hockey-analytics/

Sans, C., Carlsson, N., Lambrix, P, 2019. Player impact measures for scoring in ice hockey, In
Karlis, D., Ntzoufras, I., Drikos, S. (eds), Proceedings of MathSport International 2019 Conference,
pp. 307-317

Sarangam, Ajay 2022. Difference Between Analysis And Analytics | Which Is Better? Search date:
8.9.2022. https://www.jigsawacademy.com/blogs/business-analytics/analysis-vs-analytics/

88
Schuckers, Michael E. & Argeris, Steven 2015 . You can beat the “market”: Estimating the ROI on
NHL scouting. Journal of Sports Analytics 1 (2015) 111–119. Search date: 4.9.2022.
https://content.iospress.com/download/journal-of-sports-analytics/jsa0015?id=journal-of-sports-
analytics/jsa0015

Schulte, Oliver & Zhao, Zeyu 2017. Apples-to-Apples: Clustering and Ranking NHL Players Using
Location Information and Scoring Impact. School of Computing Science, Simon Fraser
University. Search date: 4.9.2022. https://www.cs.sfu.ca/~oschulte/files/pubs/sloan-fix.pdf

Schuckers, Michael & Curro, James. 2013. Total Hockey Rating (THoR): A comprehensive
statistical rating of National Hockey League forwards and defensemen based upon all on-ice
events. https://www.statsportsconsulting.com/wp-
content/uploads/Schuckers_Curro_MIT_Sloan_THoR.pdf Lawrence University, Iowa State
University, Statistical Sports Consultancy LLC.

SportContract. Search date: 22.2.2023. https://sportcontract.net/#services

SportContract API documentation. Search date: Multiple between November 2022 - February 2023.
https://api.sportcontract.net/api-docs/

Sudipto Guha, Nina Mishra, Gourav Roy,Okke Schrijvers. 2016. Robust Random Cut Forest Based
Anomaly Detection On Streams. Amazon.

Thomas, G, 2018. Hockey Analytics: Ditching Plus Minus. Search date: 10.10.2022.
https://puckprose.com/2018/07/30/hockey-analytics-ditching-plus-minus/

Tran, Bill 2021. How analytics and big data are changing ice-hockey. Search date: 3.9.2022.
https://thewincolumn.ca/2021/09/20/how-analytics-and-big-data-are-changing-ice-hockey/

Turtoro CJ, 2014. Catch-all Statistics Part I: GVT versus Point Shares. Search date: 12.10.2022.
https://www.allaboutthejersey.com/2014/8/8/5978425/catch-all-statistics-part-i-gvt-versus-point-
shares

89
Vik, Jon & Shih, Min-Chun & Jansher, Rabnawaz & Carlsson, Niklas & Lambrix, Patrick. 2021. Not
all goals are equally important - a study for the NHL. Linkoping University. Search date:
6.9.2022.https://www.researchgate.net/publication/285020018_A_Markov_Model_for_Hockey_M
anpower_Differential_and_Win_Probability_Added

Vollman, Rob & Fyffe, Iain & Awad, Tom. 2016 The Ultimate Guide to Hockey Analytics. p. 13-19.
Hockey Abstract. ECW Press.

What is Random Cut Forest? Amazon. Search date: 2.3.2023.

https://docs.aws.amazon.com/quicksight/latest/user/what-is-random-cut-forest.html

Wikipedia. PDO. Search date: 4.1.2023. https://en.wikipedia.org/wiki/Analytics_(ice_hockey)#PDO

Wikipedia. Liiga. Search date: 25.2.2023.

https://fi.wikipedia.org/wiki/J%C3%A4%C3%A4kiekon_SM-liiga

Wikipedia. HockeyAllSvenskan. Search date: 25.2.2023.

https://en.wikipedia.org/wiki/HockeyAllsvenskan

Wikipedia. Data Scraping. Search date: 25.2.2023. https://en.wikipedia.org/wiki/Data_scraping

Wisehockey API documentation, Wisehockey API Automated real-time statistics for third party
services 2021-09-16 V.2.1, Bitwise Oy.

Yost, Travis. 2012. Player Analysis: Goals Versus Threshold. Search date: 11.10.2022.
https://www.hockeybuzz.com/blog/Travis-Yost/Player-Analysis-Goals-Versus-
Threshold/134/45704

90
APPENDICES
Table and view structure appendix 1
Jupyter notebook for data extraction appendix 2

91
TABLE AND VIEW STRUCTURE APPENDIX 1

Liiga_2021_analysis_view_v1_1
1 playerid int
2 lastname string
3 firstname string
4 jersey int
5 role string
6 home/away string
7 matchid int
8 timeonice double
9 plus int
10 minus int
11 plusminus int
12 corsifor int
13 cosrsiagainst int
14 corsifor% double
15 pdo double
16 team string
17 expectedgoals double
18 goals bigint
19 date date

2021allplayersandteams
1 playerid int
2 firstname string
3 lastname string
4 jersey int
5 role string
6 teamid int
7 teamshortname string

2021allshots
1 team string
2 period bigint
3 result string
4 shooter bigint
5 expectedgoals double
6 teamstrength.type string
7 matchid bigint

2021allteams
1 teamid int
2 teamname string
3 teamshortname string

2021playerstats
1 numberofmatchesplayed int
2 playerid int
3 skatingstatistics.topspeed double
4 skatingstatistics.timeonice double
5 skatingstatistics.distancetravelled double
6 skatingstatistics.averagespeedwithpuck double

92
7 skatingstatistics.accelerations int
8 skatingstatistics.decelerations int
9 faceoffstatistics.faceoffwins int
10 faceoffstatistics.faceoffcount int
11 plusminusstatistics.plus int
12 plusminusstatistics.minus int
13 plusminusstatistics.total int
14 shotstatistics.shots int
15 shotstatistics.goals int
16 passstatistics.successfulpasses int
17 passstatistics.allpasses int
18 passstatistics.receivedpasses int
19 passstatistics.totalpassdistance double
20 passstatistics.forwardpassdistance double

v_player_team_count
1 teams played bigint
2 playerid int
3 lastname string

2021allmatch
1 matchid int
2 matchdate string
3 status string
4 homegoals int
5 awaygoals int
6 hometeam.id int
7 hometeam.fullname string
8 hometeam.shortname string
9 awayteam.id int
10 awayteam.fullname string
11 awayteam.shortname string
12 venue.name string
13 venue.city string

2021allmatchplayerstat
1 playerid int
2 homeaway string
3 totalstatistics.skatingstatistics.topspeed double
4 totalstatistics.skatingstatistics.timeonice double
5 totalstatistics.skatingstatistics.distancetravelled double
6 totalstatistics.skatingstatistics.averagespeedwithpuck double
7 totalstatistics.skatingstatistics.accelerations int
8 totalstatistics.skatingstatistics.decelerations int
9 totalstatistics.shiftstatisticssummary.shifts int
10 totalstatistics.shiftstatisticssummary.averageshiftduration double
11 totalstatistics.passstatistics.successfulpasses int
12 totalstatistics.passstatistics.allpasses int
13 totalstatistics.passstatistics.receivedpasses int
14 totalstatistics.passstatistics.totalpassdistance double
15 totalstatistics.passstatistics.forwardpassdistance double
16 totalstatistics.plusminusstatistics.plus int
17 totalstatistics.plusminusstatistics.minus int
18 totalstatistics.plusminusstatistics.total int
19 totalstatistics.puckcontrolstatistics.puckcontroltime double

93
20 totalstatistics.puckcontrolstatistics.puckcontroldistance double
21 totalstatistics.puckcontrolstatistics.puckcontrolforwarddistance double
22 totalstatistics.puckconteststatistics.puckcontestswon int
23 totalstatistics.puckconteststatistics.puckcontestslost int
24 totalstatistics.puckconteststatistics.puckcontestparticipations int
25 totalstatistics.shotscreenstatistics.offensivescreens int
26 totalstatistics.shotscreenstatistics.blockedshots int
27 totalstatistics.traditionalstatistics.corsifor int
28 totalstatistics.traditionalstatistics.corsiagainst int
29 totalstatistics.traditionalstatistics.corsiforpercentage double
30 totalstatistics.traditionalstatistics.relativecorsiforpercentage double
31 totalstatistics.traditionalstatistics.fenwickforpercentage double
32 totalstatistics.traditionalstatistics.relativefenwickforpercentage double
33 totalstatistics.traditionalstatistics.pdo double
34 totalstatistics.traditionalstatistics.fullstrengthsavepercentage double
35 totalstatistics.traditionalstatistics.fullstrengthshootingpercentage double
36 matchid int

2021allplayers
1 playerid int
2 firstname string
3 lastname string
4 jersey int
5 role string
6 teamid int

liiga2021allgamesgoals
1 playerid bigint
2 goals bigint
3 matchid bigint
4 lastname string
5 firstname string

rs2021playerswithteamsxg
1 playerid bigint
2 lastname string
3 firstname string
4 xg double
5 matchid bigint
6 jersey int
7 role string
8 teamshortname string

runkosarja2021allshotsxg
1 shooter bigint
2 xg double
3 matchid bigint

sc-liiga2021-skaters
1 id string
2 firstname string
3 lastname string
4 position string
5 teamid string

94
6 teamname string
7 jerseynumber double
8 gp double
9 g double
10 a double
11 p double
12 pm double
13 plusminus double
14 pp double
15 sh double
16 as double
17 toipergame double
18 fow double
19 fol double
20 fo_perc double
21 bls double
22 sog double
23 a1 double
24 a2 double
25 ppg double
26 shg double

sc-allsvenskan2021-skaters
1 id string
2 firstname string
3 lastname string
4 position string
5 teamid string
6 teamname string
7 jerseynumber double
8 gp double
9 g double
10 a double
11 p double
12 pm double
13 plusminus double
14 pp double
15 sh double
16 as double
17 toipergame double
18 fow double
19 fol double
20 fo_perc double
21 bls double
22 sog double
23 a1 double
24 a2 double
25 ppg double
26 shg double

95
JUPYTER NOTEBOOK FOR DATA EXTRACTION APPENDIX 2

import requests
from requests.structures import CaseInsensitiveDict
headers = CaseInsensitiveDict()
headers["Accept"] = "application/json"
headers["X-API-KEY"]= "xxxxxxx"
import requests
import json

url_leagues = "https://api.sportcontract.net/leagues"
print (url_leagues)
resp_leagues = requests.get(url_leagues, headers=headers).json()
print(resp_leagues)

resp
type(resp_leagues)
from pandas import json_normalize
df_league=json_normalize(resp_leagues)

df_league
import requests
from requests.structures import CaseInsensitiveDict
headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
parameters["leagueId"] = "5f4fa8e5e933ac773ac51bfd"
parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["X-API-KEY"]= "xxxx"
import requests
url_players = "https://api.sportcontract.net/players"
print (url_players)
print (parameters)
print (headers)
resp_players = requests.get(url_players, params=parameters, headers=headers).json()
from pandas import json_normalize
df_players=json_normalize(resp_players)
df_players.head()
import requests
from requests.structures import CaseInsensitiveDict
headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
#parameters["leagueId"] = "5f4fa8e5e933ac773ac51bfd"
parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["X-API-KEY"]= "xxxx"

import requests
url_games = "https://api.sportcontract.net/leagues/5f4fa8e5e933ac773ac51bfd/games"
print (url_games)
print (parameters)
print (headers)
resp_games = requests.get(url_games, params=parameters, headers=headers).json()
from pandas import json_normalize
df_games=json_normalize(resp_games)
df_games.head()

print(df_games.iloc[0, 0])

96
import requests
from requests.structures import CaseInsensitiveDict
headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
parameters["leagueId"] = "5f4fa8e5e933ac773ac51bfd"
parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["X-API-KEY"]= "xxxx"

import requests
url_game =
"https://api.sportcontract.net/games/55c8d10fa7a84b636d94ad61;55c8d10fa7a84b636d94ad57;2022-05-
04"
print (url_game)
print (parameters)
print (headers)
resp_game = requests.get(url_game, params=parameters, headers=headers).json()
from pandas import json_normalize
df_game=json_normalize(resp_game)
df_game#.head()

df_game.to_csv('//xxx/My Drive/OAMK/Thesis/code/game_Example.csv')
df_game.to_json('//xxx/My Drive/OAMK/Thesis/code/game_Example.json')

import requests
from requests.structures import CaseInsensitiveDict
headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
parameters["leagueId"] = "5f4fa8e5e933ac773ac51bfd"
parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["X-API-KEY"]= "xxxx"
import requests
url_skater= "https://api.sportcontract.net/playerstats/skater"
print (url_skater)
print (parameters)
print (headers)
resp_skater = requests.get(url_skater, params=parameters, headers=headers).json()
from pandas import json_normalize
df_skater=json_normalize(resp_skater)
df_skater.head(10)
df_skater.to_csv('//xxx/My Drive/OAMK/Thesis/code/skater_allsvenskan_2021.csv')
df_skater.to_json('//xxx/My Drive/OAMK/Thesis/code/skater_allsvenskan_2021.json')
import requests
from requests.structures import CaseInsensitiveDict
import requests

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
parameters["leagueId"] = "5f4fa8e5e933ac773ac51bfd"
parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["X-API-KEY"]= " xxxx"

url_skater= "https://api.sportcontract.net/playerstats/skater"
print (url_skater)
print (parameters)
print (headers)
resp_skater = requests.get(url_skater, params=parameters, headers=headers).json()
from pandas import json_normalize

97
df_skater=json_normalize(resp_skater)
df_skater.head(10)
df_skater.to_csv('//xxx/My Drive/OAMK/Thesis/code/allsvenskan/skater_hasvn_2021.csv')
df_skater.to_json('//xxx/My Drive/OAMK/Thesis/code/allsvenskan/skater_hasvn_2021.json')

# LIIGA and Wisehockey from this point on

from matplotlib.font_manager import json_dump
import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import requests
import matplotlib

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= " xxxx"

url= "https://api.wisehockey.com/v4.1/tournaments"
print (parameters)
print (headers)
url_resp= requests.get(url, headers=headers).json()
print(url_resp)
from pandas import json_normalize
url_resp
df_tournaments=pd.DataFrame(url_resp)

df_tournaments.to_csv('//xxx/My Drive/OAMK/Thesis/code/tournaments_wisehockey.csv')
df_tournaments.to_json('//xxx/My Drive/OAMK/Thesis/code/tournaments_wisehockey.json')

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
#parameters["leagueId"] = "55c8c0c9a7a84b636d8843e1"
#parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxxx"

url= "https://api.wisehockey.com/v4.1/tournaments"
print (parameters)
print (headers)
url_resp= requests.get(url, headers=headers).json()
print(url_resp)
from pandas import json_normalize
url_resp
df_tournaments=pd.DataFrame(url_resp)

print (df_tournaments)

df_tournaments.to_csv('//xxx/My Drive/OAMK/Thesis/code/tournaments_wisehockey.csv')
df_tournaments.to_json('//xxx/My Drive/OAMK/Thesis/code/tournaments_wisehockey.json')
#fetch all team for tournament id 31
#from urllib.parse import SplitResult
#from cv2 import split
f#rom sqlalchemy import false

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
headers["Accept"] = "application/json"

98
headers["Wisehockey-Api-Key"]= "xxxx"

url= "https://api.wisehockey.com/v4.1/tournaments/31/teams"
url_resp= requests.get(url, headers=headers).json()
url_resp
df_teams=pd.DataFrame(url_resp)

#print (df_teams)
#df_teams.to_csv('//xxx/My Drive/OAMK/Thesis/code/skaters/teams.csv')
#df_teams.to_json('//xxx/My Drive/OAMK/Thesis/code/skaters/teams.json')

jj22= (url_resp['teams'])
print (jj22)
df22=pd.DataFrame(jj22)
print (df22)
df22.to_csv('//xxx/My Drive/OAMK/Thesis/code/skaters/teams.csv', index=False)
df22.to_json('//xxx/My Drive/OAMK/Thesis/code/skaters/teams.json', orient='split', index=False,
force_ascii=False)

extracted_teamids= df22['id'].to_list()
print (extracted_teamids)

#this block will extract all players within teamid (Kalpa) for tournament id 31 =2021-2022 runkosarja

from unittest import registerResult

from matplotlib.font_manager import json_dump
import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import requests
import matplotlib as pyplot

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()

headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxx"

teamids =['859884935']
results = []

for teamid in teamids:

results = []
try:
res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/teams/{}/players'.format(teamid),
headers=headers).json()
print(res)

filename=teamid

df=pd.DataFrame(res)

except:
print ('Exception with playerid '+teamid)

#this block will extract all players with all teams for tournament id 31 =2021-2022 runkosarja

99
from base64 import encode
from distutils.log import info
from unittest import registerResult
from matplotlib.font_manager import json_dump
import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import requests
import matplotlib as pyplot

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()

headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxxx"

teamids = extracted_teamids
#teamids=['495643563','859884935','55786244']
#results = []
df=pd.DataFrame([])
df_players=pd.DataFrame([])

for teamid in teamids:

try:
res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/teams/{}/players'.format(teamid),
headers=headers).json()
#print(res)
#filename=teamid
#new_data=new_data.append(res)
print (teamid)
jj22= (res['players'])
df=pd.json_normalize(jj22)
df_players=df_players.append(df, ignore_index=True)

except:
print ('Exception with teamid '+teamid)

print (df_players)

df_players.to_csv('//xxx/My Drive/OAMK/Thesis/code/skaters/allplayers.csv', index=False)

df_players.to_json('//xxx/My Drive/OAMK/Thesis/code/skaters/allplayers.json', orient= 'split' ,
force_ascii=False, index=False)

extracted_playerids= df_players['id'].to_list()
print (extracted_playerids)
len(extracted_playerids)
#this and next block will extract all players in all teams
from unittest import registerResult
from matplotlib.font_manager import json_dump
import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import requests

100
import matplotlib as pyplot

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()
#parameters["leagueId"] = "55c8c0c9a7a84b636d8843e1"
#parameters["season"] = "2021"
headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxxx"

#playerids =['31046431','60591153', '27471447', '1234567']

playerids =extracted_playerids
results = []

df2=pd.DataFrame([])
for playerid in playerids:
try:
res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/players/{}/skatersummary'.format(playerid
), headers=headers).json()
#data=json.loads(res.text)
y = {"playerid":playerid}
res.update(y)
filename=playerid
print (filename)
df=pd.json_normalize(res)
df2=df2.append(df)

with open('//xxx/My Drive/OAMK/Thesis/code/skaters/'+filename+'.json', 'w') as f:

json.dump(res, f)

#df['PlayerId']=playerid
#print (df)

#json_out=df.to_json()
#df.to_json("//xxx/My Drive/OAMK/Thesis/code/skaters/"+filename+"_out.json", index=False,
force_ascii=False, orient="split")
df2.to_csv("//xxx/My Drive/OAMK/Thesis/code/skaters/allplayerstats_out.csv", index=False)
except:
print ('Exception with playerid '+playerid)
#Get match ids
#https://api.wisehockey.com/v4.1/tournaments/31/matches?includeMatchesWithoutStatistics=false

from matplotlib.font_manager import json_dump

import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import matplotlib

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()

headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxx"

df_matches=pd.DataFrame([])

try:

101
 res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/matches?includeMatchesWithoutStatistics=
false', headers=headers).json()
jj22= (res['matches'])
df_matches=pd.json_normalize(jj22)
except:
print ('Exception in getting matches.')

len(df_matches)

df_matches.to_csv('//xxx/My Drive/OAMK/Thesis/code/skaters/allmatches2021.csv', index=False)

df_matches.to_json('//xxx/My Drive/OAMK/Thesis/code/skaters/allmatches2021.json', orient= 'split' ,
force_ascii=False, index=False)
extracted_matchids= df_matches['id'].to_list()
#print (extracted_matchids)
len(extracted_matchids)

matchid_chunks = [extracted_matchids[x:x+50] for x in range(0, len(extracted_matchids), 50)]

#print (matchid_chunks[0])

print (len(matchid_chunks))

matchids=(matchid_chunks[0])

#print (matchids)
#get all games - playerstatistics
#https://api.wisehockey.com/v4.1/tournaments/31/matches/219244634/skaterstatistics

from matplotlib.font_manager import json_dump

import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import matplotlib

headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()

headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxx"

df_skaterstatspergame=pd.DataFrame([])
df2_skaterstat = pd.DataFrame([])
#matchids =extracted_matchids
#matchids= ['448', '367', '445', '446', '449', '450', '447', '444', '442', '364', '440', '439', '443', '441', '437',
'438', '382', '48803', '315', '422', '420', '48802', '48800', '293', '324']
matchids=(matchid_chunks[0])
counter='-0'

for matchid in matchids:

try:
res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/matches/{}/skaterstatistics'.format(matchid
), headers=headers).json()
#data=json.loads(res.text)
y = {"matchid":matchid}
res.update(y)
filename=matchid

102
 print (filename)
df_skaterstatspergame=pd.json_normalize(res['skaterStatistics'])
df_skaterstatspergame['matchid']=matchid
df2_skaterstat=df2_skaterstat.append(df_skaterstatspergame)
df2_skaterstat=df2_skaterstat.drop(columns=['periodStatistics', 'totalStatistics.speedZoneStatistics'])
with open('//xxx/My Drive/OAMK/Thesis/code/skaters/matchstats/'+filename+'.json', 'w') as f:
json.dump(res, f)

except:
print ('Exception with matchid '+matchid)

df2_skaterstat.to_csv("//xxx/My
Drive/OAMK/Thesis/code/skaters/matchstats/allmatch2021"+counter+"_out.csv", index=False)

import pandas as pd
import glob
import os

# setting the path for joining multiple files

files = os.path.join("//xxx/My Drive/OAMK/Thesis/code/skaters/matchstats/", "allmatch2021*.csv")

# list of merged files returned

files = glob.glob(files)

print("Resultant CSV after joining all CSV files at a particular location...");

# joining files with concat and read_csv

df_concat = pd.concat(map(pd.read_csv, files), ignore_index=True)
print(df_concat)

df_concat.to_csv("//xxx/My
Drive/OAMK/Thesis/code/skaters/matchstats/concat/allmatch2021_concat_out.csv", index=False)
#get all games - shotstatistics - including xG per shot!!!
#block 1
#https://api.wisehockey.com/v4.1/tournaments/31/matches/444/goals?normalize=true

from matplotlib.font_manager import json_dump

import requests
import json
import pandas as pd
from requests.structures import CaseInsensitiveDict
import matplotlib

#block 2
headers = CaseInsensitiveDict()
parameters = CaseInsensitiveDict()

headers["Accept"] = "application/json"
headers["Wisehockey-Api-Key"]= "xxx"

df_shotlstats=pd.DataFrame([])
df2_shotlstats = pd.DataFrame([])
#matchids =extracted_matchids
#matchids= ['448', '367', '445', '446', '449', '450', '447', '444', '442', '364', '440', '439', '443', '441', '437',
'438', '382', '48803', '315', '422', '420', '48802', '48800', '293', '324']
#matchids= ['448', '367']

103
matchids=(matchid_chunks[0])
counter='-0'
#block 3
for matchid in matchids:
try:
res =
requests.get('https://api.wisehockey.com/v4.1/tournaments/31/matches/{}/shots?normalize=true'.format(
matchid), headers=headers).json()
#data=json.loads(res.text)
y = {"matchid":matchid}
res.update(y)
filename=matchid
print (filename)
df_shotlstats=pd.json_normalize(res['shots'])
df_shotlstats['matchid']=matchid
df2_shotlstats=df2_shotlstats.append(df_shotlstats)
df2_shotlstats=df2_shotlstats.drop(columns=['startXPosition',
'startYPosition','secondsFromPeriodStart','speed','blocker','saver','screeningPlayers',
'shotAreaId','fromLateralPass', 'royalRoadCrossed', 'shooterSpeed','shotDirection.horizontal',
'shotDirection.vertical', 'shotDirection.goalHorizontalLimits', 'shotDirection.goalVerticalLimits',
'shotDirection'])
with open('//xxx/My Drive/OAMK/Thesis/code/skaters/matchstats/shotstats/'+filename+'shots.json',
'w') as f:
json.dump(res, f)

except:
print ('Exception with matchid '+matchid)

df2_shotlstats.to_csv("//xxx/My
Drive/OAMK/Thesis/code/skaters/matchstats/shotstats/allmatch2021"+counter+"shots_out.csv",
index=False)

import pandas as pd
import glob
import os

# setting the path for joining multiple files

files = os.path.join("//xxx/My Drive/OAMK/Thesis/code/skaters/matchstats/shotstats/",
"allmatch2021*.csv")

# list of merged files returned

files = glob.glob(files)

print("Resultant CSV after joining all CSV files at a particular location...");

# joining files with concat and read_csv

df_concat = pd.concat(map(pd.read_csv, files), ignore_index=True)
print(df_concat)

df_concat.to_csv("//xxx/My
Drive/OAMK/Thesis/code/skaters/matchstats/shotstats/concat/allmatch2021shots_concat_out.csv",
index=False)

104