FORECASTING – Part 1

 Introduction
 Forecasting methods
 Time series methods
 Stationary time series
 Forecast errors
 Time Series with a Trend

Sources: [2] Ch.2 and [3] Ch.12

1
 Introduction: What is forecasting?
A forecast is a prediction of what will occur in the future.
Forecasting the weather
Forecasting the inflation rate
Forecasting the winners of football games
Forecasting the demand of a product
A «forecast of demand» is the basis for most important
planning decisions.
Planning decisions regarding:
 production planning,
 inventory,
 facility layout and design,
 workforce,
 purchasing, and so on.

2 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Characteristics of forecasts
 They are usually wrong.
 A good forecast is more than a single number (both a range and a
measure of accuracy).
 Aggregate forecasts are more accurate (with lower s/m).
 compare weekly demand of N(100,900) vs. monthly demand of
N(400, 4*900); monthly demand forecasts are more accurate than
weekly demand forecasts.
 Forecast for sweatshirts of all sizes is more accurate than the forecast
of M-size only.
 The longer the forecast horizon, the less accurate the forecast will
be.
 compare weekly demand of the next week N(100,900) vs. weekly
demand in the next six months (additional impacts in the meantime)
N(100, 900 + ??).
 exchange rate of next Friday (Nov. 17) vs. exchange rate on Friday
four weeks ahead (Dec. 8).

3 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Characteristics of forecasts
 In general, the forecast error is higher in the upstream stages of the
supply chain.
 Forecast of sweatshirts at the factory level is less accurate than the
forecast at the distributors level.
 Forecasts should not be used without considering subjective
methods.
 Forecasts should not be used to the exclusion of known information.

4 IE 323 Lecture Notes - 9 13-15/11/2023

Introduction: Effect of inaccurate forecasting
on the supply chain

5 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Forecast horizon (tahmin ufku)
Forecast horizon indicates how far into the future is forecast.
Planning decisions rely on:
accurate short-term forecasts of sales, usually on a weekly basis;
also long-term forecasts for 12-24 months into the future.
Short-range to mid-range forecast
typically encompasses the immediate future
daily, weekly, or monthly forecasts, for up to two years as the horizon
Ex: at HP monthly forecasts for printers are made for an horizon of 12
to 18 months into the future.
Long-range forecast
usually the forecast horizon encompasses a period of time longer
than two years; and required for strategic planning.
HP’s long-term forecasts are made for years 2 through 6. At Fiat,
strategic plans for new and continuing products are for 10 years.
6 IE 323 Lecture Notes - 9 13-15/11/2023
 Introduction: Demand behaviour
Demand sometimes behaves in a random, irregular way.
Random variations
movements in demand that do not follow a pattern, and without
«assignable» causes.
At other times it exhibits predictable behaviour with trends
or repetitive patterns  trends, cycles, and seasonal
patterns.
Trend
a gradual, long-term ‘up’ or ‘down’ movement of demand.

7 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Demand behaviour
Cycle
an up-and-down repetitive movement in demand over a lengthy
time span (more than a year).
For ex: Demand for winter sports equipment increases every 4
years before the Winter Olympics.

8 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Demand behaviour
Seasonal pattern
an up-and-down repetitive movement in demand occurring
periodically in the short run. Often weather-related.
For ex: Every winter the demand for snowblowers and skis
increases.
A seasonal pattern may occur on a daily or weekly basis.
For ex: some restaurants are busier during lunch time than during
dinner time  seasonal pattern during the day

9 IE 323 Lecture Notes - 9 13-15/11/2023

 Introduction: Demand behaviour
Demand behaviour frequently displays several of these
characteristics simultaneously.
For example, demand for skis is seasonal; however there
has been an upward trend in the demand for winter sports
equipment in the past two decades.

10 IE 323 Lecture Notes - 9 13-15/11/2023

 Forecasting methods
Qualitative – Based solely on judgement
Use management judgment, expertise, opinion, past
experience, or best guesses, to make forecasts
Management, marketing, purchasing, and engineering are
sources for internal qualitative forecasts.
 Sales Force Composites: a forecasting method wherein the sales agents
forecast the sales in their respective territories, which is then consolidated
at branch/region/area level, after which the aggregate of all these factors
is consolidated to develop an overall company sales forecast.
 Customer Surveys – qualitative data (web surveys)
 Delphi Method: A procedure for acquiring informed judgment and
opinions from knowledgeable individuals using a series of questionnaires
to develop a consensus forecast about what will happen in the future.
 involves soliciting forecasts about technological advances from experts
(especially for technological forecasting).
11 IE 323 Lecture Notes - 9 13-15/11/2023
 Forecasting methods
Quantitative – Based solely on quantitative data and the
assumption that past patterns are true for the future
Mathematical methods
Time Series Methods: statistical techniques that use historical
demand data to predict future demand
Regression (or causal) Methods: attempt to develop a
mathematical relationship between demand and factors that
cause its behavior based on mathematical formulas

12 IE 323 Lecture Notes - 9 13-15/11/2023

 Steps of the forecasting process

1. Identify the 2. Collect historical 3. Plot data and identify

purpose of forecast. data. patterns.

6. Check forecast 5. Develop/compute 4. Select a forecast

accuracy with one or forecasts for the period model that seems
more measures. of historical data. appropriate for data.

7.
Is accuracy of No 8b. Select a new
forecast forecast model or
acceptable? adjust parameters of
the existing model.
Yes
9. Adjust the forecast 10. Monitor results
8a. Forecast over based on additional and measure forecast
planning horizon. qualitative information and accuracy.
insight.

13 IE 323 Lecture Notes - 9 13-15/11/2023

 Time series methods
 Time Series
Assume that what has occurred in the past will continue to occur
in the future  makes use of historical data
Relate the forecast to only one factor  time
Include the methods as naïve forecast, moving average,
exponential smoothing, and linear trend line
 Naïve forecast
demand in current period t, Dt, is used as next period’s (t+1)
forecast, Ft+1  Ft+1 = Dt.
 Simple moving average
uses average demand for a fixed sequence of periods
good for stable demand with no pronounced behavioral patterns
 Weighted moving average
higher weights are assigned to most recent data (as in
exponential smoothing methods)
14 IE 323 Lecture Notes - 9 13-15/11/2023
 Time series
A collection of observations at discrete points in time
{Dt, t ≥ 0} like {Sales week6, Sales week7, Sales week8, ..}
A time series of forecasts
{Ft , t ≥ 0} : Forecast of period t made in period t-1
(one-step-ahead-forecast)
General expression for the one-step-ahead forecast:

Ft   a n Dt n  Ft -1, t
n 1

{Ft, t+ , t ≥ 0} : Forecast of period t+ made in period t

(multiple step-ahead-forecast)
15 IE 323 Lecture Notes - 9 13-15/11/2023
 True Time Series (realizations or actual observations)
 Actual observations: {Dt, t ≥ 0}
 Never known but assumed to follow some pattern (örüntü),
mixed with random deviations (i.e., error)
 Realizations (actual observations) are
«a deterministic value from the true pattern (unknown, so
to be estimated) + random error»
 Random error t is assumed to be normally distributed:
t N(0, 2) with unknown 2
 t ’s are not assumed statistically independent, i.e.,
Cov(t , t+1 )0.
t (and hence  ) are estimated by the forecast error time
series {et, t ≥ 0} where et = Ft – Dt
16 IE 323 Lecture Notes - 9 13-15/11/2023
True Time Series (realizations or actual observations)

17 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
Assumed pattern: Dt =  + t
• Purely random
• Stationary in some medium term (longer than
the forecast horizon)

Forecasting methods:
• Moving average, MA(N)
• Exponential Smoothing, ES()

18 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
Moving Average of order N: MA(N)
The forecast is the arithmetic average of the most recent N
observations:
 N
1
Ft   an Dt n Ft   Dt n
n 1 n 1 N

Exponential Smoothing with the smoothing constant a : ES(a)

The forecast is the weighted average of the last forecast and
the current value of demand:

Ft  α Dt 1  (1  α) Ft 1
where 0<a ≤ 1 is the smoothing constant.
19 IE 323 Lecture Notes - 9 13-15/11/2023
 Stationary Time Series
Ft  Ft 1  α (Ft 1  Dt 1 )  Ft 1  α et -1
 correcting the latest forecast, Ft-1 by the realized error.

Note that:
 
Ft   an Dt n Ft   α (1  α) n 1
Dt n ,
n 1 n 1

and coefficients (weights of observations, ai) add up to 1



  α (1  α)
i 0
i
1

20 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
The weights can be approximated with an exponential
function f(i)= a e –ai. This is why this method is called
exponential smoothing (i is the age of the data point).
For example, for Ft, Dt-1 has an age of zero, as seen below:

Period

21 IE 323 Lecture Notes - 9 Age 2 1 0 13-15/11/2023

 Stationary Time Series
Example
Sales
120

t Dt 100

1 19
80
2 15
3 102 60

4 90 40

5 39 20

0
1 2 3 4 5

A stationary time series pattern looks appropriate (no trend or

seasonality can be justified in this time series).

22 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
Example
One-step Ahead Forecasts:

MA(5) ES(0.1)
t Dt Ft et Initialize F t et
1 19
2 15
3 102
4 90
5 39 53 14.0
6 53.0 51.6

23 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
Example
One-step Ahead Forecasts:

MA(5) ES(0.1)
t Dt Ft et Initialize F t et
1 19
2 15
3 102
4 90
5 39 53 14.0
6 29 53.0 24.0 51.6 22.6
7 55.0 49.3

24 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
Example: One-step Ahead Forecasts
MA(5) ES(0.1)
t Dt Ft et Initialize F t et
1 19
2 15
3 102
4 90
5 39 53 14.0
6 29 53.0 51.6 22.6
7 90 55.0 -35.0 49.3 -40.7
8 46 70.0 24.0 53.4 7.4
9 30 58.8 28.8 52.7 22.7
IE 323 Lecture Notes - 9

10 66 46.8 -19.2 50.4 -15.6

11 80 52.2 -27.8 52.0 -28.0
12 89 62.4 -26.6 54.8 -34.2

13-15/11/2023
13 82 62.2 -19.8 58.2 -23.8
14 17 69.4 52.4 60.6 43.6
15 26 66.8 40.8 56.2 30.2
25 16 29 58.8 29.8 53.2 24.2
 Stationary Time Series
Example: One-step Ahead Forecasts
MA(5) ES(0.1)
t Dt Ft et Initialize F t et
1 19
2 15
3 102
4 90
5 39 53 14.0
6 29 53.0 51.6 22.6
7 90 55.0 -35.0 49.3 -40.7
8 46 70.0 24.0 53.4 7.4
9 30 58.8 28.8 52.7 22.7
IE 323 Lecture Notes - 9

10 66 46.8 -19.2 50.4 -15.6

11 80 52.2 -27.8 52.0 -28.0
12 89 62.4 -26.6 54.8 -34.2
13 82 62.2 -19.8 58.2 -23.8
14 17 69.4 52.4 60.6 43.6
15 26 66.8 40.8 56.2 30.2
26
16 29 58.8 29.8 53.2 24.2 13-15/11/2023
 Stationary Time Series
Example: One-step Ahead Forecasts

27 IE 323 Lecture Notes - 9 13-15/11/2023

 Stationary Time Series
MA and ES parameters
The larger the N in MA(N)
Less emphasis on recent realizations (less responsive)
More stable forecasts
The larger the  in ES()
More emphasis on recent realizations (more responsive)
Less stable forecasts
Standard deviation of et , that is se, in terms of s of each
demand observation (Appendix 2-A in [2] textbook):
N 1 to equate  e ' s in the two methods :
σe  σ in MA(N)
N 2
N -1
2 α
σe  σ in ES(α) 2
2α α
28
IE 323 Lecture Notes - 9
N 1 13-15/11/2023
 Stationary Time Series
 Ft  1, t  τ  Ft for 
Both methods are poor in detecting trends and they lag:

29 IE 323 Lecture Notes - 9 13-15/11/2023

 Testing by studying forecast errors

How to evaluate the accuracy of forecasts?

1 n
Mean Absolute Deviation (MAD)   ei
n i 1
1 n 2
Mean Square Error (MSE)   ei
n i 1
1 n ei
Mean Absolute Percentage Error (MAPE)   * 100 %
n i 1 Di
If forecast error distribution is N(0,e2) in truth, then
e  1.25 MAD.
It is also desired not to have error bias, i.e., E(ei) = 0
30 IE 323 Lecture Notes - 9 13-15/11/2023
 Bias in Forecasts

31 IE 323 Lecture Notes - 9 13-15/11/2023

 Testing: Forecast Errors
Example

32 IE 323 Lecture Notes - 9 13-15/11/2023

 Time Series with a Trend
Assumed pattern
D t =  +  t + t
• Non-stationary
Forecasting methods
• Double Exponential Smoothing, DES(, ) - Holt’s method
Estimate Dt and δ.
St : Base forecast level (Dt estimate)
Gt : Slope estimate (δ estimate)
Ft,t τ  St  τ Gt , τ  1
St  α Dt  (1  α)(St 1  Gt 1 )
Gt  β(St  St 1 )  (1  β)Gt 1
33 IE 323 Lecture Notes - 9 13-15/11/2023
Time Series with a Trend
St  α Dt  (1  α)(St 1  Gt 1 )
Gt  β(St  St 1 )  (1  β)Gt 1 Forecast,
(Ft+1)
Ft,t τ  St  τ Gt , τ  1
Dt Gt

St

Realizations

Dt-1
Forecast, St - St-1
(Ft) Gt-1
Base
Level
St-1

t-1 t t+1
Time periods
34 IE 323 Lecture Notes - 9
13-15/11/2023
Time Series with a Trend
Single and Double Exponential Smoothing are alike.
In ES() forecast for period t

Ft  α Dt -1  (1  α)Ft 1
In DES(, ) forecast for period t+1

St  α Dt  (1  α)(St 1  Gt 1 )

Gt  β(St  St 1 )  (1  β)Gt 1

Ft 1  St  Gt

35 IE 323 Lecture Notes - 9 13-15/11/2023

Time Series with a Trend
Initialization:
Good practice to initialize by linear regression
Fit a line to a series of historical data points
Minimize the sum of the squares of the deviations of the
actual observations from the projected line Dt = a + bt
where Dt is the dependent variable (variable to be
predicted  demand), and t is the independent variable
(time).

36 IE 323 Lecture Notes - 9 13-15/11/2023

Time Series with a Trend
 Initialization: example (using some initial data for initialization)

37 IE 323 Lecture Notes - 9 13-15/11/2023

Time Series with a Trend

Cov(Dt , t ) Cov(Dt , t )
b 
Cov(t,t ) Var (t )

1 n 1 n n n
1 n n

n t 1
tDt  2  t  Dt
n t 1 t 1
 tDt   t  Dt
n t 1 t 1
 n n
 t 1
n n
1 1 1

n t 1
t 2
 2 
(
n t 1
t ) 2

t 1
t 2
 (
n t 1
t ) 2

1 n n1
a   Dt  b
n t 1 2
38 IE 323 Lecture Notes - 9 13-15/11/2023
Time Series with a Trend: example
Year Period Index Sales Sales
1969 1 4675 10000
1970 2 5289 9000
1971 3 5811 8000
7000
1972 4 6294
6000 Sales
1973 5 7083 5000
1974 6 6552 4000
1975 7 6942 3000
2000
1976 8 7842
1000
1977 9 8514 0
1978 10 9269 1 2 3 4 5 6 7 8 9 10

Year Period Index Sales t2 t Dt

1969 1 4675 1 4675
1970 2 5289 4 10578
1971 3 5811 9 17433 Cov(Dt,t) Var(t)
1972 4 6294 16 25176 3774.45 9.167
1973 5 7083 25 35415 b a
1974 6 6552 36 39312 411.76 4562
1975 7 6942 49 48594
1976 8 7842 64 62736
1977 9 8514 81 76626
1978 10 9269 100 92690 a+ bt 8680
39 SUM 55 68271 385 413235 13-15/11/2023
IE 323 Lecture Notes - 9
Time Series with a Trend: example
Holt’s Method Applied
DES(0.15, 0.1)
From LR Initialization
a b
0.15 0.1
t Dt St-1+Gt-1 St Gt
9 8268 412
1978 10 9269 8680 8768 421

a b
0.15 0.1
t Dt St-1+Gt-1 St Gt
9 8268 412
1978 10 9269 8680 8768 421
1979 11 9189

40 IE 323 Lecture Notes - 9 13-15/11/2023

Time Series with a Trend: example
Holt’s Method Applied
DES(0.15, 0.1)
a b
0,15 0,10
t Dt St-1+Gt-1 St Gt
9 8268 412
1978 10 9269 8680 8768 421
1979 11 9636 9189 9256 427
1980 12 9683
a b
0,15 0,10
t Dt St-1+Gt-1 St Gt
9 8268 412
1978 10 9269 8680 8768 421
1979 11 9636 9189 9256 427
1980 12 11043 9683 9887 448
1981 13 10335
41 IE 323 Lecture Notes - 9 13-15/11/2023