Received 7 March 2023, accepted 20 March 2023, date of publication 29 March 2023, date of current version 5 April 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3263260

Real-Time Portfolio Management System

Utilizing Machine Learning Techniques
PRAKASH K. AITHAL 1, (Member, IEEE), GEETHA M. 1 , (Member, IEEE),
U. DINESH ACHARYA 1, BASRI SAVITHA2 , AND PARTHIV MENON 1
1 Department of Computer Science and Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal 576104, India
2 Manipal Institute of Management, Manipal Academy of Higher Education, Manipal 576104, India
Corresponding author: U. Dinesh Acharya ([email protected])

ABSTRACT There are 1641 companies listed on the National Stock Exchange of India. It is undoubtedly
infeasible for a retail investor to invest in all the stocks. It is a well-known fact that the portfolio’s return is
an average return of all its constituent stocks, and risk will be less than or equal to the maximum risk of all
the portfolio components. This paper is unique as it elaborates on the entire portfolio selection, optimization,
and management process. Portfolio selection is accomplished through the K-Means algorithm. Optimization
is achieved utilizing the genetic algorithm, and a sliding window is applied for portfolio management.
Four different ways of portfolio calculation, namely, equally-weighted portfolio, global minimum variance
portfolio, market cap-weighted portfolio, and maximum Sharpe ratio portfolio, are applied. The results depict
that all three optimized portfolios outperform the Nifty index. The dataset for the study is obtained from
globaldatafeeds.in.

INDEX TERMS Portfolio selection, portfolio optimization, portfolio management, real-time, K-means
algorithm, metaheuristic algorithms, maximum sharpe ratio portfolio, global minimum variance portfolio,
equally-weighted portfolio, sliding window.

I. INTRODUCTION tinuously readjusting the funds allocated to the stocks to gain

People invest in assets in the hope that they will bring future maximum profit at minimal risk [1].
benefits. There are asset classes: treasury bills, bonds, stocks, The portfolio is managed in two different ways, namely,
gold, real estate, etc. People will invest in a combination of passive portfolio management and active portfolio manage-
these asset classes. The present study considers stocks only. ment. In passive portfolio management, fund managers either
A stock portfolio is a collection of stocks. The advantage of invest their client’s money in index funds or selected stocks.
a portfolio is its return is the weighted return of its constituent The built passive portfolio will not alter till its maturity. Pas-
stocks. The portfolio’s risk will be less than the risks of its sive portfolio managers assume that the markets are efficient.
constituent stocks. Contrary to passive portfolio management, active portfolio
The portfolio selection is an NP-hard problem. Once it is management involves investing in stocks and readjusting the
decided how much money needs to be invested in stocks, the portfolio whenever there is a change in market conditions.
portfolio selection process begins. Active portfolio managers assume that there is a chance of
Portfolio selection has two steps: The choice of stocks is making a profit whenever markets deviate from efficient
made in the first step. The second step decides how much form. Active portfolio managers do fundamental analyses,
money goes into selected stocks. Portfolio optimization is event studies, and technical analyses of the stock market [2].
carried over to adjust the money that is allocated to each There are numerous stocks traded every day. Every investor
chosen stock. Portfolio management is the process of con- wishes to maximize their profit and minimize their risk.
In Pre 1950s, people used to invest in individual stocks and
The associate editor coordinating the review of this manuscript and were exposed to the risk of the instruments they invested in.
approving it for publication was Vlad Diaconita . Markowitz proved that risk could be reduced by investing
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME 11, 2023 32595
 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

in a set of instruments known as a portfolio. The portfo- The stop-loss and take-profit are required to tackle extreme
lio gives the average return of constituent instruments with profit or losses. The genetic algorithm is capable of identify-
reduced risk. Portfolio selection, optimization, and manage- ing stop-loss and take-profit points [14]. A genetic algorithm
ment became pioneer research fields. The computer-aided was utilized to track the investment. The portfolio will move
finance made the field even more interesting as it increased in which direction was studied [15]. A Jordan Elman network
the stakeholder base, reduced the transaction cost, and so with a genetic algorithm was used for portfolio selection. The
on [3]. method uncovers the hidden patterns in the stock market [16]
The portfolio selection, optimization, and management Investor sentiment is fed to the grouping genetic algorithm
are helpful for institutional investors as they can build and to optimize the portfolio. The Chen et al. have arrived at
manage their portfolios. The proposed manuscript will help following rules [17]
retail investors to build and manage their portfolios. The
1) Rule 1-On a day t, if a stock’s Buy Sell Imbalance is
proposed work can be extended to any nation’s stock market,
Low and DayTrading is rising, then buy the stock.
and investors can make informed decisions on investing in
2) Rule 2- On a day t, if a stock’s Buy Sell Imbalance is
the stock market. Fund managers can offer ETFs based on
High and DayTrading is rising, then sell the stock.
the proposed manuscript.
Mohammad Maholi Solin et al. have proposed an artificial
neural network model to predict the stock prices and a genetic
II. LITERATURE REVIEW algorithm to optimize the portfolio. Their study reveals that
The modern portfolio theory put forth by Harry Markowitz the genetic algorithm outperforms the single-index model.
in the early 1950s gave insight that instead of investing in a The study involved 38 stocks from the Indonesian stock mar-
single stock, it is profitable to invest in a set of stocks known ket: the data considered open, close, high, low, and volume
as a portfolio. Experts from several domains have investigated [18]. An electricity market portfolio is established through
the construction and management of portfolios. The decision the metaheuristic genetic algorithm. Metaheuristic genetic
to buy, sell or hold the stocks in the portfolio is a tedious task. algorithm achieves similar results compared to the exact algo-
To select the instruments for the portfolio and to overcome rithm but with higher efficiency [19]. Gupta et al. categorize
the sector bias, the hierarchical clustering algorithm is used. the stocks into three groups using a support vector machine.
The clustering is required to get the maximum returns from From each class, stocks are picked to form the portfolio uti-
a portfolio. The clustering-based algorithm’s volatilities are lizing the genetic algorithm [20]. For uncertain environments,
much smaller than the Markowitz Mean-Variance model [4], one can not rely upon mean and variance alone. Skewness
[5], [6]. The primary goal of any investor is to make a and Kurtosis need to be considered as well. In an uncertain
profit. Diversification and proper allocation of wealth are environment, it is not sufficient to depend upon historical
necessary to achieve the goal. Fund managers do it manually data; expert opinion must be taken into account [21]. Stock
for their clients. A computer-based portfolio selection was performance depends on economic features such as produc-
presented [7]. It is impossible to manually analyze the stock tion power, service sector, industrialization, farming, natural
movements and decide about buying, selling, or holding the resources, etc. Finance must reflect the current economy [22].
stock. Portfolio optimization is usually accomplished through A genetic algorithm-based portfolio optimization technique
a genetic algorithm [8].There are three operators in a genetic was utilized to optimize the SPX500 stock market, and it
algorithm: crossover, mutation, and inversion. A grouping outperforms the market by a significant margin [23].
genetic algorithm is applied by Chen et al. [9] for finding The literature review shows that the metaheuristic genetic
a group trading strategy portfolio. A group stock portfolio algorithm provides highly accurate results with greater effi-
categorizes the stocks into different groups: banking stocks, ciency. The genetic algorithm can be applied for portfolio
IT stocks, Health stocks, and so on. A grouping genetic selection and optimization.
algorithm is applied to optimize the portfolio. If any stock To avoid overfitting by genetic algorithm sliding window is
within the sector is not performing, it is replaced with stock applied. The training window’s length must be proportionate
from the same sector [10]. Chen et al. have proposed an to the size of the testing window, say 12:1, 6:1, or 3:1 [24]
improvement to their earlier article by implementing the A multi-objective portfolio selection model using max-
fuzzy grouping genetic algorithm [11]. A genetic algorithm imum entropy with a genetic algorithm was presented.
is a nature-inspired algorithm. The genetic algorithm iterates The model achieved higher returns with more diversifi-
over the population and finds the optimal solution. It applies cation due to the introduction of maximum entropy [25].
crossover, mutation, and selection operations to the popu- The benefit of including the real estate investment trust in
lation to find the optimal solution [12]. The genetic algo- the portfolio was studied using a genetic algorithm. The
rithm and Particle Swarm Optimization algorithm are two genetic algorithm outperformed the global minimum vari-
renowned nature-inspired meta-heuristic algorithms. They ance portfolio [26]. Diversification-aware portfolios were
reduce computational time and can manage realistic con- optimized using a genetic algorithm-Out of the three port-
straints efficiently [13]. The trading strategy involves stop- folios constructed, One of the portfolios underperformed.
loss, take-profit points, and long-term portfolio management. One performed marginally better, and one outperformed the

32596 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

benchmark [27]. A survey of different portfolio optimiza- assets. This results in the creation of an equal weight index
tion methods was presented [28]. The stock price prediction for each of the assets. Equal weight indices are also referred
application utilizing metaheuristics algorithms is an area yet to as unweighted indices.
to be explored [29]. The support of parallelism in genetic The performance of each asset in an equally weighted
algorithms, simplicity, and efficiency makes it more suitable portfolio is responsible for the performance of the entire
for portfolio optimization problems [30]. portfolio. The comparatively greater performance of equally-
Behavioral finance is an interdisciplinary field involv- weighted portfolios can be attributed in part to its strategy of
ing psychology, event studies, economics, and sociol- value investing. Value investing involves picking those stocks
ogy. Humans make their investment decisions based on for the portfolio that appear to be undervalued and have the
emotions [31]. potential to grow in the future. Such assets, more often than
The reviewed literature does not cover all the phases of not, give higher returns over a longer period.
portfolio management. Only one metaheuristic algorithm was Another defining feature of equally-weighted portfolios
considered to optimize the portfolio in the reviewed litera- is their diversification. Diversification is a risk management
ture. The validation of obtained results was missing in the strategy that combines a variety of assets in the portfolio.
literature. It works on the principle that having a mixture of different
The proposed work covers the entire portfolio manage- types of assets will lower the risk posed to one particular asset
ment. Seven metaheuristic algorithms are compared. The and yield higher returns over time.
constructed portfolio is validated. The manuscript presents all Equally weighted portfolios also feature a higher portfolio
the phases of the portfolio, namely, portfolio selection, port- turnover rate. Turnover rate is defined as the ratio in which
folio optimization, and portfolio management. The holistic assets in a fund are bought and sold by portfolio managers,
approach is not present in the reviewed literature. which ultimately gives the percentage change of assets over
a period and the consequent higher transaction costs and
III. METHODOLOGY taxes. Additionally, when the market experiences prolonged
The methodology is presented in subsections III-A and III-B declines during the bear market phase, equally-weighted port-
with titles Models and Process, respectively. folios are more vulnerable to sudden changes in value [32].

2) MARKET CAP WEIGHTED PORTFOLIO

A. MODELS
A market capitalization-weighted portfolio, often shortened
The following are the key concepts and models used in this
to a market-cap-weighted portfolio, consists of a portfolio
article.
where individual assets are included in values that corre-
1) Equally-Weighted (EW) Portfolio spond to their total market cap. The individual components
2) Market Cap Weighted (MCW) Portfolio of the portfolio are calculated by multiplying the outstanding
3) Global Minimum Variance (GMV) Portfolio shares (owned by individual shareholders, institutional block
4) Maximum Sharpe Ratio (MSR) Portfolio holdings, and company insider holdings) of a company by
5) K-Means clustering Algorithm the current price of a single share, which corresponds to the
6) Genetic Algorithm (GA) current market price of the outstanding share.
7) Ant Colony Optimization (ACO) Algorithm The total value of such a portfolio keeps varying based on
8) Particle Swarm Optimization (PSO) Algorithm the price of the share in the market as well as the weight
9) Firefly Algorithm allocated to a particular asset in the portfolio. Market capital-
10) Artificial Bee Colony (ABC) Optimization Algorithm ization is a representation of the total market value of a firm’s
11) Tabu Search outstanding shares and the contribution of its shareholders.
12) Simulated Annealing (SA) In the market-cap-weighted portfolio, assets with a higher
13) Sliding Window. market cap will receive a higher percentage of the weight
from the portfolio. Companies that fare comparatively poor
1) EQUALLY-WEIGHTED PORTFOLIO in the market receive a lesser percentage of the weight from
Equal weight is a proportional measuring method that assigns the total portfolio. This ensures that poorly performing stocks
the same importance to every stock in a portfolio, index, do not affect the overall performance of the overall index.
or index fund. In such a system, stocks of smaller and lesser- A big advantage of this type of index is that it gives access
known companies are given the same importance as stocks to a wide spectrum of companies. Even though this makes
of larger and more important ones. An index with equally it vulnerable to frequent stock movements, the presence of
weighted assets must continuously buy and sell as shares of larger companies might mean that the share values of such
companies increase and decrease in price. Shares are bought companies can give constant growth in value to the portfo-
when the prices of a security start to decline, and balance is lio. Additionally, some companies have shares that are not
restored where as shares are sold when the prices increases. completely public, forcing indices to use the technique of free
Therefore, an equally weighted portfolio of assets implies float factor (the percentage of shares available for trading) for
that the capital invested is distributed equally among all the calculating prices and fluctuations.

VOLUME 11, 2023 32597

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

In contrast, large companies having higher weights can Having a higher Sharpe ratio, generally above 1.0, indi-
have a disproportionate impact on the portfolio as a whole. cates a better risk-adjusted performance. If portfolio analysis
Market-cap-weighted portfolio, hence, gives an almost unbi- results in a negative Sharpe ratio, then it implies either that
ased automatic process for promoting stocks with higher the risk-free rate is greater than the return of the portfolio or
values and negating ones with declining prices [32]. that the return of the portfolio is actually negative in value.
On similar lines, a well diversified portfolio would have
3) GLOBAL MINIMUM VARIANCE PORTFOLIO a higher sharpe ratio compared to other portfolios having
Even with the existence of a variety of optimized portfolios, a lower level of diversification. Sharpe ratio can hence be
the returns that can be expected from a particular portfolio used as a tool to identify and compare values of funds in
can be difficult to determine in most cases. Mathematically, two different categories. Shares that give similar returns but
the global minimum portfolio is the most efficient portfolio have different risk factors involved can be identified using
since it minimizes risks and maximizes returns from the this method.
investment. This occurs because the weights of this portfolio One of the limitations of using the Sharpe ratio as a mea-
depend only on the return variance values and the covariances sure for analyzing risk-adjusted performance is that it relies
and not on the returns from the portfolio. on standard deviation as a critical parameter of calculation.
On a broader note, the global minimum variance portfolio It just assumes that returns are normally distributed. This
is a subset of the efficient frontier representation. The effi- may lead to erroneous results, faulty investments, and lesser
cient frontier is a set of optimal portfolios that give the best returns over a longer time period [32].
returns with the least risk, for any given level of return.
Portfolios lying below the efficient frontier are sub-optimal 5) K-MEANS CLUSTERING
portfolios as they do not give enough returns for that level of
Unsupervised learning algorithms analyze and create clusters
risk, whereas those to the right are sub-optimal because they
in unlabeled datasets. These algorithms are widely used for
have a higher level of risk involved. Successfully optimizing
exploratory data analysis and image recognition because of
a portfolio would imply that the portfolio will lie on the
their ability to detect and group patterns on their own.
efficient frontier line.
Clustering is an unsupervised data mining technique that
The global minimum variance portfolio, hence, would lie
groups data points into separate clusters, where the data
to the leftmost end of the efficient frontier. It represents a
points in a particular group have similar properties. More
portfolio having the least risk, by virtue of least standard
specifically, exclusive clustering, also known as hard cluster-
deviation, but still provides optimal returns to the investor.
ing, implies that one particular data point can be part of only
Stocks in any minimum variance portfolio might be risky
one particular cluster. The K Means clustering algorithm is
when held individually but balance each other out when held
one such simple and well-known exclusive clustering algo-
together under the same index.
rithm. Additionally, K Means clustering is also an example
Hence, diversification of investment is an apt concept to be
of partition clustering as the entire data is divided into non-
applied to minimum variance portfolios in general. Investing
hierarchical partitions by virtue of centroids.
in sectors with little correlation can lead to a less volatile
In the K Means algorithm, all the data points are organized
portfolio, thus giving larger returns over a more extended
into K groups, where K represents the number of centroids
period [32].
present in the dataset. A centroid is an imaginary location
depicting the center of the entire cluster. Every data point is
4) MAXIMUM SHARPE RATIO PORTFOLIO
then assigned to the nearest groups while at the same time
The Sharpe ratio is used to help investors evaluate the return keeping the centroids as small as possible.
on investment compared to its risk. The ratio defines the The means in K Means refers to the averaging of data to
average return in addition to the free risk rate upon the total find the centroid for every cluster. A larger K value in the
risk involved. algorithm indicates smaller groupings and more granularity,
The Sharpe ratio is calculated by subtracting the risk-free while a smaller K value indicates larger groups with less
rate from the return of the portfolio and then dividing the granularity.
result by the standard deviation of the portfolio. The standard The algorithm starts its execution with a randomly selected
deviation is a measure of the volatility of the portfolio. The set of centroids. These centroids are used as starting points
Sharpe ratio hence is an efficient methodology to calculate for the respective clusters. Further iterative calculations are
risk-adjusted returns on an index. performed on the dataset to optimize the positions of these
The formula to calculate the Sharpe ratio is given in centroids such that either of the following conditions is
equation 1. met:
(R(p) −R(f ))/s(p) (1)
• The positions of the centroids stop changing because
where R(p) indicates the return of the portfolio, R(f) indicates they have been optimized
the risk-free rate of return, and s(p) denotes the standard • The predetermined number of iterations has been com-
deviation of the portfolio pleted. This leads to a successful clustering operation.

32598 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

Even with all the advantages of K Means clustering, there 7) ANT COLONY OPTIMIZATION ALGORITHM
are a few drawbacks to it. They are not suitable for noisy Ant Colony Optimization (ACO) is classified under ant
data and cannot identify non-convex clusters. The number of colony algorithms and applies specific metaheuristic opti-
clusters to be formed also needs to be given in advance to the mizations as well. A probabilistic technique that reduces
algorithm. Additionally, since it is an unsupervised learning computational problems to finding the shortest path between
algorithm, it takes a longer training time, has a higher risk graphs, it takes inspiration from the collective foraging
factor of producing inaccurate results, and is ambiguous in behavior of ant colonies. It was first introduced by Marco
the way it classifies clusters [33]. Dorigo in 1992 [35] as a means to search for an optimal
path in a graph. Eventually, it began to be applied to more
numerical applications.
6) GENETIC PROGRAMMING The crux of the algorithm is to observe and replicate the
Genetic Programming is a type of evolutionary algorithm movement of ants from their nests in search of food. Ants tend
which is a subset of Machine Learning. They are a part to take the shortest path available to the food resource. Thus,
of heuristic search algorithms that are based on the prin- they end up walking over the shorter path more often than
ciple of natural selection and evolution. They are utilized on other paths. The presence and persistence of pheromones,
to solve problems that humans cannot solve directly. Since chemicals that ants use to communicate, on these paths help
they are free from human intervention, these algorithms and determine the length of the path followed. The more ants
programming techniques are used in optimization problems march over a particular trail, the higher the pheromone den-
and search algorithms. sity and the better the optimization. This mimicry of simu-
A genetic program consists of generations that, in turn, lated ants finding the shortest path and leaving behind artifi-
consist of a population of individuals. Each individual entity cial pheromones by walking around a graph characterizes the
represents a point in the search space. Each individual is ACO algorithm.
represented by a string of characters, bits, or floats.
Cleverly exploiting random search, the genetic program-
8) PARTICLE SWARM OPTIMIZATION ALGORITHM
ming methodology generates a set of solutions which are
Particle Swarm Optimization (PSO) is a population-based
in turn represented by sets of fixed-length strings. Each of
stochastic technique that optimizes a given problem by trying
these strings is referred to as a chromosome. From a set of
to improve a solution concerning the quality of the solution.
best solutions, newer and improved solutions are created by
PSO, first proposed by Kennedy and Eberhart [36], was
reiterating the entire process. This process is repeated until
initially aimed at simulating the social behavior of groups of
either the best possible solution is obtained or the required
organisms like flocks of birds or a school of fish.
number of iterations is completed.
Each particle is considered part of a research space con-
Even though the algorithm appears to be random, there are
taining a global minimum. None of the particles are aware
three major operations that go into the genetic programming
of the exact location of this minimum but have a general
process.
idea of its location. Each particle is attributed with a certain
Selection involves taking chromosomes with good fitness
velocity vector that is randomly chosen. The movement of
scores from the parents and passing them on to successive
the particles is affected by cognitive influence (towards its
generations for fitter and better offspring. It is based on the
previous personal best position) and social influence (towards
idea of survival of the fittest.
the best solution of the group).
The crossover operation involves actual mating between
The coefficients of inertia, cognitive and social influence
individuals. Crossover sites are chosen randomly, and better
control the levels of exploration and exploitation, which are
strings are created by exchanging genes at these locations.
important in deciding the solution to the problem. Exploration
The mutation is an operation that keeps running in the
implies the ability of particles to move towards the best solu-
background. It facilitates a change in genes or chromosomes
tions found till a particular time. Exploitation, on the other
to allow for diversity in the population. It allows the algorithm
hand, refers to the ability of particles to process the entire
to be on the lookout for newer possibilities for the optimal
research space. Optimization using PSO often involves find-
solution.
ing the right balance between exploration and exploitation.
Genetic algorithms are more robust than traditional arti-
ficial intelligence algorithms as they do not converge or
break under the presence of noise in the input stream. 9) FIREFLY ALGORITHM
They also have a higher ability to optimize search prob- The firefly algorithm is a metaheuristic inspired by the behav-
lems. Genetic programming can be applied in the classi- ior of fireflies and their bioluminescence. It is believed that
fication of data, automated bug fixing, ensuring network lesser brighter fireflies will be attracted to brighter ones.
security by rapid identification and resolving of breaches, If there are no fireflies nearby, then a particular firefly moves
and as a supporting mechanism for various other machine randomly.
learning methods like neural networks and recurrent neural From the initial population of fireflies that exist, light
networks [34]. intensity is formulated that is related to the function in

VOLUME 11, 2023 32599

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

question. An absorption coefficient controls the brightness • Long-term memory

of the light. The attractiveness of the fireflies varies with - - It is based on the frequency of occurrence and helps
distance. As the fireflies move towards each other, newer the algorithm explore new spaces
solutions are computed, and the light intensity varies accord- - - It is handled by the frequency memory
ingly. In the end, the fireflies are ranked to find the firefly The algorithm is repeatedly invoked on a search space until a
having the best value for the objective function. particular number of iterations has been achieved or the best
This effectively subdivides the whole population into sub- solution is found [37].
groups as each subgroup swarms around a local optima [29].
12) SIMULATED ANNEALING
10) BEE COLONY OPTIMIZATION ALGORITHM
Simulated Annealing is a probabilistic optimization algo-
Bee Colony Optimization (BCO), a population-based algo- rithm used to tune model parameters so that they do not get
rithm, is inspired by the activity and movements of bees stuck at the local minimum while searching for an optimal
in nature. It belongs to the category of swarm intelligent global minimum value of a function. It draws inspiration from
algorithms, a subset of AI algorithms based on the study of the physical annealing process, where a material is heated
actions of individuals in various decentralized systems. Arti- up until it reaches an annealing temperature and is cooled
ficial bees are programmed such that when acting together, down slowly so that the required structure is obtained from
they can solve complex combinatorial problems. the material.
The algorithm has two alternating stages: the forward pass At each interaction, the algorithm probabilistically deter-
and the backward pass. In every forward pass, the bees mines whether or not to move the system to one of its neigh-
explore newer search spaces (similar to how they search for boring states. The probability of moving from the current
nectar in newer locations). Here, a predefined number of state to a new neighboring state is defined by the accep-
moves is applied that helps to generate a new partial solution tance probability function and depends on the energies of the
to the problem. Every backward pass consists of bees sharing two states as well as temperature, which varies with time.
all the information that was obtained during the forward pass The evaluation and acceptance of the neighbors from every
(similar to how bees share information about nectar locations move may result in the system temporarily accepting and
in the beehive). moving to a worse state before moving on to find the global
These artificial bees have objective function values that minimum [34].
characterize the quality of the solution. Based on a random
probability, the bee decides whether to abandon the created 13) SLIDING WINDOW
partial solution or to recruit more bees to return to the created The sliding window algorithm is used to solve problems
partial solution. It is observed that bees having a better objec- involving linear sequences. A window is formed that slides
tive function value have a higher probability of continuing over the given data, thus covering different parts of the data
its exploration. Every artificial bee generates a solution to at different points in time. The main aim behind this algorithm
the problem. The best possible solution is selected at the deals with converting two nested loops into one single loop.
end [34]. This reduces time complexity to linear time.
The sliding window technique has two basic requirements.
11) TABU SEARCH
• There must be two pointers, say, L and R, that corre-
Tabu Search employs local search methods to optimize math- spond to the left and right ends of the current working
ematical problems. Similar to a few other metaheuristics, this range
technique also considers the neighbors of a potential solution • The algorithm must be able to add new elements to the
in the hope of discovering a better solution. range when the right-most pointer is moved forward and
Tabu Search is different from other local search techniques must be able to delete or remove elements from the range
in that it allows for worse states of neighbors to be accepted. when the leftmost pointer is moved forward.
This prevents the algorithm from being stuck at a local min-
ima and helps it explore regions of the search space not Based on the size of the ranges, there are two types of slid-
visited earlier. Additionally, restrictions (or tabu) are applied ing window techniques in practice. If the length of the sliding
on regions of the search space visited earlier, thus preventing window remains fixed, it is known as the fixed-size sliding
the algorithm from considering the same solution repetitively. window technique. On the other hand, if the length of the
There are different kinds of memory structures involved in sliding window is modifiable, it is called the flexible window
tabu search. size sliding technique.
This technique of applying a sliding window is beneficial
• Short-term memory in the fact that it prevents unnecessary iterations over a collec-
- - It is based on how recent occurrences are handled tion of data. It also gives a convenient way in order to analyze
and prevents the algorithm from going to recently data in distinct sub-sections.
visited search spaces A major disadvantage of the sliding window technique
- - It is handled by a Tabu list of traversal of arrays or strings is the computational cost

32600 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

involved. Increasing the window size makes it faster but at algorithm. The following parameters are utilized in GA as
the same time affects the accuracy of operation. they gave the best results.
1) maximum number of iteration = 1000
B. PROCESS 2) population size = number of stocks
The actual stock data from globaldatafeeds.in is considered 3) mutation probability = 0.5
for the study. The data comprised of stock ticks data for 4) elit ratio = 0.01
a period of one month. The data consists of NSE NIFTY 5) crossover probability = 0.06
100 stock prices. The comprehensive list of NIFTY-100 6) parents portion = 0.3
companies is depicted in table 1. The query for storing the 7) crossover type = uniform
historical data fields is depicted in figure 2. The instidentifier 8) maximum iteration without improvement = 0.01.
field identifies the instrument and open, high, low, and close Mutation probability is the allowed probability of change
fields depicts the open price, high price, low price and close in population. Crossover probability is the probability that
price respectively. Traded quantity field stores the volume of interpopulation change occurs. Elit ratio specifies how many
trade for the particular instrument. The query for storing the entries need to be unaltered for the next generation. The uni-
real-time data fields into the timescale db is depicted in figure form crossover ensures that the genes are chosen randomly,
3. This quantitative data, along with the qualitative data, and there is no positional bias. The algorithm for portfolio
is utilized to build the portfolio. The four different portfo- optimization is given below.
lios, namely, equally weighted, market-cap-weighted, global The following parameters are utilized for the PSO.
minimum variance, and maximum Sharpe ratio, are managed 1) nostalgia = 0.8
with the help of K-Means clustering and metaheuristic algo- 2) envy = 1.3
rithms. The study considered forty-four financial indicators 3) inertia weight = 0.6.
depicted in table 2. The seven financial indicators are chosen Since PSO was designed to find the global optima, the envy
using principal component analysis. The seven significant value considered is greater than the nostalgia value. The
financial indicators, namely, the corporate tax rate, GDP, inertia weight decides the speed of the particle. To utilize
current account to GDP, import, industrial production, prime the ACO, SA, and Tabu search in the model, The model
lending rate, and tourist arrivals [38] are given as input to the was redesigned as a network. The historical portfolio returns
system. The calendar of events data [39], tone of the annual and current day portfolio return are utilized to predict the
report, stock prices, and volume are fed to the proposed future portfolio returns. In case of ACO, Euclidian distance
system as depicted in the Fig. 1. is applied to obtain the optimal solution efficiently. SA and
A stock portfolio is a collection of stocks. Portfolio selec- Tabu search required the input to be in the form of linear
tion and optimization for an active portfolio management and binary quadratic model. ABC model is trained with the
technique are proposed. The proposed portfolio construction following parameters to gain the optimal results.
and optimization are based on quantitative and qualitative
1) colony_size = 40
data. The quantitative data comprise macroeconomic indica-
2) scouts = 0.5
tors, stock prices, and stock volume. The qualitative data con-
3) iterations = 50
tains Tweets, news articles, the director’s report, and auditor
4) min_max = ‘min’
reports. The elbow method is applied for the ten-day sliding
5) nan_protection = True
window to identify the best K of the K-means algorithm.
6) log_agents = True.
The portfolio which maximizes the return is constructed. The
following optimization methods are explored As sliding window slides, the previous optimal portfolio
is compared with the current optimal portfolio, and the best
1) Genetic Algorithm among the two is retained.
2) Ant Colony Optimization Algorithm
3) Particle Swarm Optimization Algorithm IV. RESULTS
4) Firefly Algorithm The K-means clustering is utilized to group the stocks. The
5) Artificial Bee Colony Optimization Algorithm K-means algorithm with different numbers of stocks and
6) Tabu Search different values of K is considered, and an optimal K-value is
7) Simulated Annealing. found utilizing the elbow method. The elbow plot for K = 5
The genetic algorithm with the minimum variance weights, and a total of 5 stocks is depicted in figure 4. The elbow plot
and Sharpe weights objective functions are considered. For for K = 100 and a total of 100 stocks is shown in figure 5. It is
each of the objective functions, the portfolio returns are evident from the graph that as data size increases, the value of
computed. The training period of six months and one year best-K also increases. The K-means clustering with optimal
is utilized, and a testing period of one month is considered. K = 5 for datasize of 10 stocks is given in the figure 6. In the
K-means clustering partitions the N observations into K clus- figure X and Y axis are PCA axis.
ters. Each observation belongs to the cluster with the nearest To find the MSR a negtive sharp ratio is minimized.
mean. A genetic algorithm is a search-based optimization Weights of GMV portfolio is given in table 3 and weights of

VOLUME 11, 2023 32601

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

TABLE 1. NIFTY100 instruments.

FIGURE 1. Overall architecture of proposed methodology.

MSR portfolio are given in table 4. The objective function of The evolution of particles in a PSO algorithm is depicted
the GMV and MSR portfolios are depicted in figures 7 and 8 in Figure 9. From Figure 9, it is clear that the global optima
respectively. are attained as all the particles converge to value one. The

32602 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

TABLE 2. List of macroeconomic indicators.

FIGURE 3. Query to store real-time data in TimeScaleDB.

Input Stock prices, Tone

Output Optimal Portfolio
1: procedure Active Portfolio Optimization
2: while sliding window Do
3: for K = 2 to N Do
4: Construct EW, MCW, GMV and MSR portfo-
lios by chosing stocks from different K-Means cluster
5: Construct GA with two objective functions
(MV, SR weights)
6: Compare portfolio returns with respect to K
7: end for
8: Find K which maximizes the portfolio return
9: From Derived K
10: Construct GA with two objective functions (MV,
SR weights)
11: Return four optimal portfolio
12: end while
13: end procedure

FIGURE 2. Query to store historical data.

movements of Bees in a ABC algorithm are presented in10.

From Figure 10, it is clear that, though the pattern of move-
ment is different from the particles of the PSO algorithm,
the result is that all the Bees converge to value one, thus
attaining the global optima. The movement of ants in an ACO
algorithm is depicted in Figure 11. From Figure 11, it is clear
that the global optima are attained as all the ants move to value
one. The solutions of ABC algorithm is depicted in figure 12.
As can be observed from the figure 12 the ABC algorithm is
successfully identifying the patterns. FIGURE 4. Optimal K for 5 stocks.

The average and total return of individual stock of GMV

portfolio is visualized in figures 13 and 14 respectively. The
stock ticks data is fed to the built portfolio. The figure 15 Nifty-100 returns. It can be easily observed that all the opti-
depicts the real-time optimized portfolio returns against NSE mized portfolios are performing better than the Nifty-100

VOLUME 11, 2023 32603

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

TABLE 4. Maximum sharpe ratio portfolio weights.

FIGURE 5. Optimal K for 100 stocks.

FIGURE 6. Sample KMEANS result with K = 5.

TABLE 3. Global minimum portfolio weights.

FIGURE 7. Objective function for GMV portfolio.

benchmark returns. The rolling return for four years and fifty
year time periods is shown in figures 16 and 17 respectively.
The rolling return gives an unbiased view of returns over
a period. The six-month rolling return is plotted. Similarly,
a rolling Sharpe ratio for a four and fifty-year time frame FIGURE 8. Objective function for MSR portfolio.
is depicted in figures 18 and 19 respectively. Whenever the
Sharpe ratio is above one, the investment decision is wise, and
the decision is terrible when the Sharpe ratio is below zero. is illustrated in figures 20 and 21 respectively. It is evident
Finally, a rolling beta for four and fifty years time horizon from the figure that the built model is stable as beta is always

32604 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

FIGURE 9. Evolution of PSO particles. FIGURE 13. Average return of each stock under GMV portfolio.

FIGURE 10. The bee movements.

FIGURE 11. The ant movements.

FIGURE 14. Total return of each stock under GMV POrtfolio.

FIGURE 12. ABC algorithm solutions.

TABLE 5. Different portfolio returns.

zero. The weights are optimized utilizing GA and the returns

obtained with each portfolio is depicted in table 5. It is evident
from table 5 that though all four portfolios are outperforming
the Nifty-100 benchmark index, The risk to reward ratio of FIGURE 15. Real-time portfolio returns.
the four portfolios significantly differ. GMV portfolio is a
clear winner with risk to reward ratio of 0.002. The close
competitor is MSR portfolio, with risk to reward ratio of 0.5. For the dataset under consideration, GA, ABC, ACO,
The EW and CW portfolios lag with a risk-to-reward ratio of and PSO algorithms outperform Simulated Annealing, Tabu
10.97 and 11.28, respectively. search, and Firefly algorithms. GA is more efficient than the

VOLUME 11, 2023 32605

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

TABLE 6. GMV portfolio returns for various metaheuristic algorithms.

TABLE 7. MSR portfolio returns for various metaheuristic algorithms.

FIGURE 16. Rolling return for four years.

FIGURE 18. Rolling sharpe for four years.

FIGURE 17. Rolling return for fifty years.

ABC, ACO, and PSO algorithms. Simulated Annealing, Tabu

search and Firefly algorithms gave portfolio return of 20% FIGURE 19. Rolling sharpe for fifty years.
for GMV and 25.8% for MSR portfolio respectively. The
rest of the algorithms results for GMV and MSR portfolios
are depicted in Table 6 and 7 respectively. It is evident from GA outperforms other metaheuristic algorithms as there is
Table 6 and 7 GMV portfolio outperforms all other portfoilos. only risk and return to be optimized. Different metaheuristic

32606 VOLUME 11, 2023

P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

The key takeaways of this manuscript are portfolio selec-

tion and ranking with the K-Means algorithm, portfolio
optimization through metaheuristic algorithms, and portfolio
management with the help of a sliding window.
The deep neural network models can be applied. Current
work can be extended to India’s entire NSE and BSE stock
markets. The number of clusters required for a varying num-
ber of stocks and the data size can be analyzed. In the future,
the work can be extended to the stock market of other nations.

REFERENCES
[1] Z. Bodie, A. Kane, and A. J. Marcus, Investments. New York, NY, USA:
McGraw-Hill, 2009.
[2] K. Kremnitzer, ‘‘Comparing active and passive fund management in
emerging markets,’’ Senior Honors thesis, Econ. Dept., Univ. California,
Berkeley, Berkeley, CA, USA, 2012.
[3] G. Curtis, ‘‘Modern portfolio theory and behavioral finance,’’ J. Wealth
FIGURE 20. Rolling beta for four years.
Manage., vol. 7, no. 2, pp. 16–22, Jul. 2004.
[4] A. Arévalo, D. León, and G. Hernandez, ‘‘Portfolio selection based on hier-
archical clustering and inverse-variance weighting,’’ in Computational Sci-
ence ICCS 2019, J. M. F. Rodrigues, P. J. S. Cardoso, J. Monteiro, R. Lam,
V. V. Krzhizhanovskaya, M. H. Lees, J. J. Dongarra, and P. M. A. Sloot,
Eds. Cham, Switzerland: Springer, 2019, pp. 315–325.
[5] A. A. Saji, J. J. M. Joseph, and B. S. Kumar, ‘‘Portfolio construction
and testing using cluster analysis,’’ Asian J. Manage., vol. 11, no. 2,
pp. 207–212, 2020.
[6] D. León, A. Aragón, J. Sandoval, G. Hernández, A. Arévalo, and J. Niño,
‘‘Clustering algorithms for risk-adjusted portfolio construction,’’ Proc.
Comput. Sci., vol. 108, pp. 1334–1343, Jan. 2017.
[7] Y. S. Asawa, ‘‘Modern machine learning solutions for portfolio selection,’’
IEEE Eng. Manag. Rev., vol. 50, no. 1, pp. 94–112, Mar. 2022.
[8] S. Lim, M.-J. Kim, and C. W. Ahn, ‘‘A genetic algorithm (GA) approach
to the portfolio design based on market movements and asset valuations,’’
IEEE Access, vol. 8, pp. 140234–140249, 2020.
[9] C.-H. Chen, Y.-H. Chen, J. C.-W. Lin, and M.-E. Wu, ‘‘An effective
approach for obtaining a group trading strategy portfolio using grouping
genetic algorithm,’’ IEEE Access, vol. 7, pp. 7313–7325, 2019.
[10] C.-H. Chen, C.-Y. Lu, T.-P. Hong, J. C.-W. Lin, and M. Gaeta, ‘‘An effective
approach for the diverse group stock portfolio optimization using grouping
genetic algorithm,’’ IEEE Access, vol. 7, pp. 155871–155884, 2019.
[11] C.-H. Chen, B.-Y. Chiang, T.-P. Hong, D.-C. Wang, and J. C.-W. Lin,
FIGURE 21. Rolling beta for fifty years.
‘‘An approach for diverse group stock portfolio optimization using
the fuzzy grouping genetic algorithm,’’ in Intelligent Information and
algorithms will fair better if too many objective functions are Database Systems, N. T. Nguyen, D. H. Hoang, T.-P. Hong, H. Pham, and
B. Trawiński, Eds. Cham, Switzerland: Springer, 2018, pp. 510–518.
to be optimized. [12] P.-Y. Kang, I.-C. Wu, and C.-H. Hsueh, ‘‘Applying hueristic algorithms
to portfolio selection problem,’’ in Proc. Conf. Technol. Appl. Artif. Intell.
V. CONCLUSION (TAAI), Nov. 2015, pp. 323–329.
[13] D. G. Cortes, A. J. C. Jofre, and L. S. Martin, ‘‘Artificial intelligence
All the constructed portfolios outperform the Nifty-100
based method for portfolio selection,’’ in Proc. Congreso Internacional de
index. The genetically optimized global minimum variance Innovación y Tendencias en Ingeniería (CONIITI), Oct. 2018, pp. 1–4.
portfolio outshines other portfolios. The optimized maximum [14] C.-H. Chen, Y.-H. Chen, V. G. Diaz, and J. C.-W. Lin, ‘‘An intelligent
Sharpe ratio portfolio gives the highest return, but the risk is trading mechanism based on the group trading strategy portfolio to reduce
massive loss by the grouping genetic algorithm,’’ Electron. Commerce
high. Res., vol. 23, pp. 3–42, Mar. 2021.
The seven metaheuristic algorithms are explored and it [15] M. K. de Melo, R. T. N. Cardoso, and T. A. Jesus, ‘‘A genetic algorithm
is observed that GA, ABC, ACO and PSO algorithms out- for investment tracking with stochastic model predictive control,’’ in Proc.
IEEE Congr. Evol. Comput. (CEC), Jun. 2021, pp. 1543–1550.
perform SA, Tabu search, and Firefly algorithm. The GA [16] C. Kumar and M. N. Doja, ‘‘An optimizing utility for portfolio selection
algorithm is simple and efficient compared to ABC, ACO, based on optimal values computed using ANN, NSGA-II and machine
and PSO algorithms. learning technique,’’ in Proc. Int. Conf. Power Electron., Control Autom.
(ICPECA), Nov. 2019, pp. 1–5.
The number of groupings increases with the increase in
[17] C.-H. Chen, T.-P. Hong, and S.-C. Chu, ‘‘Diverse group stock portfolio
the number of stocks which implies the stocks are falling optimization based on investor sentiment index,’’ in Proc. Conf. Technol.
into different categories. The risk to reward ratio of the Appl. Artif. Intell. (TAAI), Dec. 2017, pp. 100–103.
global minimum variance portfolio is the highest indicating [18] M. M. Solin, A. Alamsyah, B. Rikumahu, and M. A. Arya Saputra, ‘‘Fore-
casting portfolio optimization using artificial neural network and genetic
an attractive investment opportunity. The built models are algorithm,’’ in Proc. 7th Int. Conf. Inf. Commun. Technol. (ICoICT),
stable as the rolling beta is zero. Jul. 2019, pp. 1–7.

VOLUME 11, 2023 32607

 P. K. Aithal et al.: Real-Time Portfolio Management System Utilizing Machine Learning Techniques

[19] R. Faia, T. Pinto, Z. Vale, J. M. Corchado, J. Soares, and F. Lezama, GEETHA M. (Member, IEEE) received the Ph.D.
‘‘Genetic algorithms for portfolio optimization with weighted sum degree from NITK Surathkal, in 2010. She is cur-
approach,’’ in Proc. IEEE Symp. Ser. Comput. Intell. (SSCI), Nov. 2018, rently a Professor with the Department of Com-
pp. 1823–1829. puter Science and Engineering, Manipal Academy
[20] P. Gupta, M. K. Mehlawat, and G. Mittal, ‘‘Asset portfolio optimiza- of Higher Education, Manipal Institute of Tech-
tion using support vector machines and real-coded genetic algorithm,’’ nology, Manipal, India. She has presented several
J. Global Optim., vol. 53, pp. 297–315, Jun. 2012. papers at national and international conferences.
[21] S. K. Mittal and N. Srivastava, ‘‘Mean-variance-skewness portfolio opti- Her work has been published in several inter-
mization under uncertain environment using improved genetic algorithm,’’
national journals. Her research interests include
Artif. Intell. Rev., vol. 54, no. 8, pp. 6011–6032, Dec. 2021.
data mining and text mining in the healthcare and
[22] B. Li, ‘‘Research on optimal portfolio of financial investment based
on genetic algorithm,’’ in Proc. Int. Conf. Econ. Manage. Model Eng. financial sectors.
(ICEMME), Dec. 2019, pp. 497–500.
[23] P. Sinha, A. Chandwani, and T. Sinha, ‘‘Algorithm of construction of opti-
mum portfolio of stocks using genetic algorithm,’’ Int. J. Syst. Assurance
Eng. Manage., vol. 6, no. 4, pp. 447–465, Dec. 2015.
[24] Y.-H. Chou, S.-Y. Kuo, and Y.-T. Lo, ‘‘Portfolio optimization based
on funds standardization and genetic algorithm,’’ IEEE Access, vol. 5,
pp. 21885–21900, 2017.
[25] A. M. Rather, V. N. Sastry, and A. Agarwal, ‘‘Portfolio selection using
maximum-entropy gain loss spread model: A GA based approach,’’ in U. DINESH ACHARYA received the Ph.D. degree
Proc. Int. Conf. Adv. Comput., Commun. Informat. (ICACCI), Sep. 2014, from the Manipal Academy of Higher Education,
pp. 400–406. Manipal, India, in 2008. He is currently a Professor
[26] F. Z. Habbab, M. Kampouridis, and A. A. Voudouris, ‘‘Optimizing mixed- with the Department of Computer Science and
asset portfolios involving REITs,’’ in Proc. IEEE Symp. Comput. Intell. for Engineering, Manipal Academy of Higher Edu-
Financial Eng. Econ. (CIFEr), May 2022, pp. 1–8. cation, Manipal Institute of Technology, Manipal.
[27] T. Onek, ‘‘A genetic algorithm model for financial asset diversification,’’ He has presented several papers at national and
in Proc. SoutheastCon, Apr. 2019, pp. 1–6.
international conferences. His work has been pub-
[28] M. K. Pareek and P. Thakkar, ‘‘Surveying stock market portfolio optimiza-
lished in several international journals. His cur-
tion techniques,’’ in Proc. 5th Nirma Univ. Int. Conf. Eng. (NUiCONE),
Nov. 2015, pp. 1–5.
rent research interests include data analytics in the
[29] P. Agrawal, H. F. Abutarboush, T. Ganesh, and A. W. Mohamed, ‘‘Meta- healthcare, agriculture, and financial sectors.
heuristic algorithms on feature selection: A survey of one decade of
research (2009–2019),’’ IEEE Access, vol. 9, pp. 26766–26791, 2021.
[30] S. Liu and C. Xiao, ‘‘Application and comparative study of optimization
algorithms in financial investment portfolio problems,’’ Mobile Inf. Syst.,
vol. 2021, pp. 1–10, Nov. 2021.
[31] A. Antony, ‘‘Behavioral finance and portfolio management: Review of
theory and literature,’’ J. Public Affairs, vol. 20, no. 2, p. e1996, May 2020.
[32] Z. Bodie, A. Kane, A. J. Marcus, and P. Mohanty, Investments, 11th ed.
India: McGraw-Hill. BASRI SAVITHA received the M.A. degree in eco-
[33] G. A. V. Pai and T. Michel, ‘‘Evolutionary optimization of constrained k- nomics and the M.B.A., H.G.Dip. (U.K.), M.Phil.,
means clustered assets for diversification in small portfolios,’’ IEEE Trans. Ph.D., and PGDHE degrees. She is currently an
Evol. Comput., vol. 13, no. 5, pp. 1030–1053, Oct. 2009. Associate Professor with the Manipal Institute of
[34] S. Luke, Essentials of Metaheuristics, vol. 2. Raleigh, NC, USA: Lulu Management, Manipal Academy of Higher Educa-
Raleigh, 2013. tion, Manipal, India. She has evaluated the impact
[35] A. Colorni, M. Dorigo, and V. Maniezzo, ‘‘Distributed optimization by
of micro health insurance schemes in Karnataka
ant colonies,’’ in Proc. 1st Eur. Conf. Artif. Life, vol. 142, Dec. 1991,
for her Ph.D. study with the National Institute of
pp. 134–142.
Technology Karnataka. She has published more
[36] R. Eberhart, ‘‘James. kennedy, particle swarm optimization,’’ in Proc.
IEEE Int. Conf. neural Netw., Aust., vol. 1948, 1942. than 22 research articles in national and interna-
[37] C. Yu, N. Lahrichi, and A. Matta, ‘‘Optimal budget allocation policy for tional journals. Her research interests include banking and insurance, and
Tabu search in stochastic simulation optimization,’’ Comput. Oper. Res., financial markets. She was awarded the Kalam Travel Grant to present
vol. 150, Feb. 2023, Art. no. 106046. her research paper from the Kalam Research Conference, University of
[38] P. K. Aithal, A. U. Dinesh, and M. Geetha, ‘‘Identifying significant Kentucky.
macroeconomic indicators for Indian stock markets,’’ IEEE Access, vol. 7,
pp. 143829–143840, 2019.
[39] P. K. Aithal, U. D. Acharya, M. Geetha, and P. Menon, ‘‘Building a
calendar of events database by analyzing financial spikes,’’ IEEE Access,
vol. 9, pp. 114192–114206, 2021.

PRAKASH K. AITHAL (Member, IEEE) is cur-

rently pursuing the Ph.D. degree with the Manipal PARTHIV MENON is currently pursuing the
Academy of Higher Education, Manipal, India. degree with the Department of Computer Sci-
He is an Assistant Professor with the Department ence and Engineering, Manipal Academy of
of Computer Science and Engineering, Manipal Higher Education, Manipal Institute of Technol-
Academy of Higher Education, Manipal Institute ogy, Manipal, India. His research interest includes
of Technology, Manipal. He has presented several data analytics.
papers at national and international conferences.
His work has been published in several interna-
tional journals. His research interest includes data
analytics in the financial domain.

32608 VOLUME 11, 2023