Forecasting methods and

demand management

1
 Plan

Objectives and definitions of forecasts

Qualitative forecasting methods
Quantitative forecasting methods
Time series
Moving average (simple, multiple, weighted)
Exponential smoothing (simple and double)
-Causal methods
Simple and multiple linear regression
Polynomial regression

2
 Reminder: Statistical concepts
A chronological series or chronicleis constituted by a sequence
ordered observations of a quantity over time. The study of
these series serve to describe, explain, control, predict phenomena
evolving over time.

The study of a chronological series { tx, t = 1,…, T} consists of dissociating the

different movements that make it up and to analyze them.

Generally, the series are handled:

With two components: trend and residual movement.
With three components: trend, seasonal movement and
residual movement.

Note: The interest of decomposing the series is, on the one hand, better.
to understand, to better describe its evolution, and on the other hand to anticipate its
evolution (based on trends and seasonal variations).

3
 Reminder: Statistical concepts

The trend, rated ftis the factor representing the evolution to

long-term greatness, and translated the general aspect of the series
example: growth of the consumption of a product

The seasonal factor, I noted strepeats at equal time intervals with

a roughly constant form. It may be due to the rhythm of the seasons or
to human factors. Its period is 12 for monthly series,
of 4 for quarterly series... If p denotes the period of the
seasonal movement: st = st + p= st + 2p The seasonal factor is
so completely determined by p seasonal coefficients: s1,…, sj,…, sp

Irregularityalso called residue movement and noted et , group

everything that has not been taken into account by the trend and the factor
seasonal. It is the result of irregular fluctuations and
unpredictable due to non-permanent disruptive factors; these
fluctuations are supposed to have low amplitude and zero average on
a small number of consecutive observations.

4
 Reminder: Statistical concepts
The decomposition of a time series with a seasonal movement can
is carried out according to three types of models:
additive model xt= ft + st+ et t = 1,…, T
In an additive model, it is assumed that the 3 components: trend, seasonal variations, and irregularities are independent of each other.
of others.

Graphically, the amplitude of variations is constant around the trend.

1eramultiplicative model form xt = ft(1+st) + ett = 1,…, T
In this model, it is assumed that seasonal variations depend on the trend.

Graphically, the amplitude of seasonal variations varies (the seasonal movement presents amplitudes
proportional to the trend
2thform of the multiplicative model xt= ft(1+ st) .(1+et ) t = 1,…, T
In this model, it is assumed that seasonal variations and irregularities depend on the trend.

5
Definition of Forecast:

Function for estimating future demand for goods and

services offered by the company, which is established either mathematically
(historical data), either intuitively (market knowledge),
by combining the two methods.

ACGPS, "Dictionary of Production and Inventory Management"

(1993)

6
General principle of the forecasting methodology

The future demand for a product is a crucial variable for

a company
Demand forecasting is very important for a
company, because it represents the knowledge of the future
sales
In order to make predictions, one must have a certain.
knowledge of the historical data of the product in question
(observations).

7
 Objectives of forecasting and management of the
request
Once the forecasts are known, we can make some
decisions such as:
The purchasing policy for raw materials and
other products
The size of the batches to be manufactured
Levels of existence in the warehouse and the stock of
security
Priority of manufacturing orders
–Etc...
Forecasts are therefore necessary for planning.
production and inventory management
8
 The characteristics of forecasting methods

They should involve little information

Be efficient
Low cost
Must be able to adapt to fluctuations in the
demand (sensitivity)

9
 The classes of forecasting methods and the cycle
the life of a product

Qualitative methods
Quantitative methods

10
 Qualitative methods

Qualitative methods are based on judgment without

no digital processing:

–Sellers' opinion
Consumer opinion (survey)
Expert Opinion
-Opinion of the executives
–…

11
 Qualitative methods

The individual method: it is based on the opinions of

people involved in the product market
Delphi Method: it seeks to achieve a consensus among experts.
group of experts
Each person assigns a probability to the different
events
We communicate these probabilities to the experts and redo it.
process until a consensus is reached
Method of historical analogy or comparison with
similar products sold in the past
Market study method
Consumer questionnaires to anticipate
market changes

12
 Quantitative methods

Time series
Moving average
Trend and seasonality
Simple and double exponential smoothing

Causal methods
Simple and multiple regression
Generalized linear models
Additive models
Artificial neural networks
–…

13
 General principle of forecasting methodology

Increasing linear trend: y=at+b

Trend calculated by the method
least squares adapted to
historical records

From the measurements taken over a known period, we determine

a trend (mathematical function) whose shape can vary in
history function.
Then, an extrapolation is possible over a period of
forecast. The latest date of the forecasting period is called
forecast horizon HP (corresponds to a time limit and to-
beyond which the prediction is less reliable). 14
 Calculation of a trend
Least Squares Method

y=ax+b (linear model) (x is the period index)

The method consists of determining the fixed coefficients a and b.
The method of least squares minimizes the sum of the squares of the deviations.
between historical values and the points of the line (between observations and
models).
The algebraic sum of the deviations is zero.
Calculation formulas for coefficients:
• = −
• = −
.
•a =
• = −

15
 Calculation of a trend
Method of Least Squares
Example:

x=t=1 x=t=2
x or t Year 2015 2016 2017 2018 2019 2020
y Value 200 240 220 270 250 290

Calculate the forecast for the following year (2021) using the trend by
least squares method
X1=1-3,5 (21/6)=-2,5
Y1=200-245=-45
a= 15,14
b=245-(15.14*3.5)=192.01
y = 15.14 x + 192.01
y2021=15,14 *7 + 192,01=298

16
 The moving average
•Method adapted to stable requests without trend that
consists of taking the arithmetic mean of the last n
years to establish the forecast (generally 3≤n≤6)

The moving average uses the most recent observations.

to calculate the forecast.

generally we need to keep a large number of

data in memory.

From a set of observed values, we calculate their

average and we use the average as a forecast of the
next period

17
 The moving average

This method gives equal weights to each of the

"N" last values of the series, and a weight equal to zero
to the values observed before.

Each new forecast is based on an average

mobile is an adjustment of the previous average
mobile.

The smoothing effect increases as 'N' increases.

(much weaker adjustment of a forecast to
the other)

18
 The moving average

We transform a graph defined from individual values Vi into a

graph of average values Vmi calculated from a grouping of
minus 3 consecutive individual values.
The points are calculated as follows (in the case of an average with 3
values):

•Vm1=(Vi1 + Vi2 + Vi3)/3

Vm2 = (Vi2 + Vi3 + Vi4) / 3
Vm3 = (Vi3 + Vi4 + Vi5) / 3

19
 The moving average
Example:

20
 The weighted moving average

This method consists of giving more importance to values.

récentes. Ainsi, des coefficients ou poids sont attribués aux données
and the sum of the coefficients is equal to 1 (100%).

Vm1 = aVi1 + bVi2 + cVi3

Vm2=aVi2 + bVi3 + cVi4

•Vm3=aVi3 + bVi4 + cVi5

Where: a+b+c=1

21
 The weighted moving average

Example:
The calculation of the weighted average is carried out over 3 periods (months)

by applying the coefficients 0.1, 0.2, 0.7 (10%, 20%, and 70%).

Period P1 P2 P3 P4
coefficient 10% 20% 70%
Consumption VI 90 80 100
Average Vm 95

22
 The centered moving average
Calculation of centered moving averages of length p:
1
•First case, odd p, p = 2m + l : =− = +
There are (T–p + 1) centered moving averages of odd length p.
Second case, p even, p = 2m:
Mp(t) = 1/p (xt-m/2 + −1 + x / 2 )
=− +1 t+m +

23
 The centered moving average

Example: Calculation of M2, M3and M4

M2 M3 M4

24
 Mobile Ratio Method
(Seasonal sales case)

In the case of seasonal sales, this method consists of

calculate a seasonal coefficient SC for each month of
the year based on the data from the study period.
We use CS to estimate sales for the following year.

25
 Mobile report method
(Seasonal sales case)

In the case of seasonal sales, this method consists of

calculate a seasonal coefficient CS for each month of
the year based on the data from the studied period.
CS is used to estimate the sales of the following year.

CS=Consumption/Average

26
 Mobile ratio method
(Seasonal sales case)

Example:
It is a company where the average sales were 350 per month during
In the year 2020, the quantities sold amounted to 490 in January 2020 and
to 280 in July 2020.
Calculate the sales estimate for January 2021 and July 2021 knowing that
The average sales estimate is projected to be 420 per month in 2021.
Solution:
The CSJanuarySalesJanuary/moy2020
490/350=1.4
VentesJanuary 2021=CSJanuaryMy2021=1,4*420=588

The CSJulySalesJuly/moy2020
=
SalesJul2021=CSJulyMy2021=

27
 Mobile ratio method
(Seasonal sales case)
Exercise
Quarter (year sales CS Quarterly forecasts
2020 (year 2021)
1 320 320/420 = 0.76 450*0.76 = 343
2 400 400/420 = 0.95 450*0,95= 428
3 580 580/420 = 1,38 450*1.38 = 623
4 380 380/420 = 0.90 450*0.9 = 406
Total=1680

The estimate of annual sales for the year 2021 is: 1800

The average quarterly value = 450

28
 Mobile ratio method
(Seasonal sales cases)

The companies did not experience seasonal sales but

regular, the coefficient for each month would be equal to 1.

If the seasonal coefficient is > 1, the activity of the quarter or

of the month is considered to be above average.

If the coefficient is < 1, the activity of the quarter or the month is

below average.

29
 Exercise
The following table provides the bi-monthly time series of transportation of
travelers on the Air France International network (in billion passenger-km)
from 2002 to 2005.

1. Calculate the trend using the least squares method

2. Determine the trend of this chronicle by the sequence of averages.
mobiles of suitable length
3. Calculate the seasonal coefficients.

30
 Simple exponential smoothing

The exponential smoothing methods, developed by R. G.

Brown in the 60s are methods of extrapolation
which give predominant weight to recent values. They
are also characterized by the simplicity of calculations and the
small amount of data to keep in memory.

Its principle is the calculation of an average value by considering

the importance given to the observations decreases
with their seniority.

This forecasting method applies to time series without

seasonal variations and locally constant trends.

31
 Simple Exponential Smoothing

Let x1, x2,…, x nthe n first observations of a series

temporally, exponential smoothing aims to
to estimate the value of xn+1 not yet observed. We
let's note this forecast x̂n 1

Exponential smoothing defines this forecast as follows:

n1
j
xˆn1 (1 )x n j
j 0

with is the smoothing parameter; 0.1

32
 Simple exponential smoothing
Y't+1 = Y't+ α (Yt-Y’t)
With
-Y't+1the forecast for the period t+1
-Y'tthe forecast for the period t
-Ytthe observed value of period t
-α: the smoothing coefficient

33
 Simple exponential smoothing
A smoothing constant α close to 0 (≤0.3) gives a
significant importance to remote observations, while
that an α close to 1 (≥0.7) tends to neglect these observations
distant.

34
Example

35
 Exercise
The demand for a certain item has been recorded over 15 consecutive months:

1. Apply simple exponential smoothing to this time series by

taking α = 0.6 up to the 6th month inclusive and α = 0.3 for the months
following.

36