Faculty Introduction

 Name/Qualification/Experience: Tohid Kachwala, B.E.

(Mechanical), MAM, PhD (Management), Industry 1984 - 1997,
Teaching – Part Time: 1992 - 1997, Full Time: 1997 onwards
(Teaching at NMIMS since 1992)
 Present Status: Professor of Operations and Data Sciences at
SBM, NMIMS University (8th Floor, SBM Faculty Area – Cabin
Number: E851)
 Email: t.kachwala@sbm.nmims.edu
 Subjects: – Statistical Inference for Decision Making (Trim I),
Modelling and Optimization for Business Decisions (Trim II),
Production & Operations Management (Trim III), Service
Operations Management (Trim VI)
 Forecasting Methods
(Time Series Methods)

Dr. T. T. Kachwala
Importance of Forecasting

In management situations, forecasting is important because the lead

time for decision making ranges from several years (for case of capital
investment) to few days (for transportation schedules) to few hours
(for production schedules).

The important component of forecasting is the distinction between

uncontrollable external events (originating with National economy,
governments, customers and competitors) and controllable internal
events (such as marketing or manufacturing decision within the
firms). The success of a company depends on both the type of events,
but forecasting applies directly to the former (uncontrollable external
events), while decision making applies directly to the latter
(controllable internal events). Planning is the link that integrates them.
Importance / Definition of Forecasting

Suppose we forecast the sales volume for a particular

product for the next quarter of the year. Production
schedules, raw material purchasing plan, inventory
policies & sales quota will be affected by the quarterly
forecast we provide. Consequently, poor forecast will
result in increased cost for the firm.

Definition: The word forecast means projection of past

into the future for e.g. Demand forecasting: based on past
demand, we forecast future demand values.
Types of Forecasting Techniques

Two general types of forecasting techniques are used

for demand forecasting.
1. Quantitative Method
2. Qualitative Method

Quantitative forecasting methods can be used when:

1. past information about variable being forecast is available
2. information can be quantified
3. assuming that pattern of past will continue into the future
Time Series Methods

1. If historical data are restricted to past values of the

variable that we are trying to forecast, the forecasting
procedure is called time series method.
2. The objective of time series method is to discover a
pattern in the historical data and then extrapolate this
pattern into the future
3. The forecast is based solely on past values of the
variable that we are trying to forecast and past
forecast errors.
Random Fluctuation in Time Series
In time series analysis, measurements may be taken at
regular intervals. They generally exhibit random
fluctuation as indicated below (Average Sales is
Constant)
Random Fluctuations / Trend in the Time Series

Random fluctuations of the time series is the residual

factor that includes deviation of actual time series
values from those expected due to random variability in
the time series. The irregular component is caused by
the short term, unanticipated, uncontrollable and non
recurring factors that affect the time series. Because
this component accounts for random variability in the
time series it is unpredictable.

Sometimes the time series may show gradual shifts or

movements to relatively higher or lower values over a
long period of time. The gradual shifting of time series
is referred as the trend in the time series.
Trend in the Time Series

This shifting or trend is usually the result of long term

factors such as changes in population, changes in
technology & customer preferences.
Seasonal Component in Time Series
Whereas a trend component of a time series are identified by
analysing multiyear movements in historical data, many time series
show a regular pattern over 1-year period. These situations in
business & economics involve period to period comparisons
Forecasting Methods of Time Series

1. Smoothing Methods
a) Moving Average
b) Weighted Moving Average
c) Exponential Smoothing
2. Trend Projection
3. Trend Projection Adjusted for Seasonal Influence
Smoothing Methods
The objective of smoothing method is to “smooth” out “the random
fluctuations” caused by the irregular component of the time series.
The popular smoothing methods are: moving average, weighted
moving average & exponential smoothing. These smoothing methods
are appropriate for a stable time series i.e. one that exhibits no
significant trend or seasonal effects. Many manufacturing
environments require forecast for thousands of items weekly or
monthly. Thus in choosing a forecasting technique simplicity & ease
of use are important criteria. Smoothing methods are easy to use &
generally provide high accuracy for short range forecast such as
forecast for the next time period.
Simple Average Method

1
f t 1   d t  d t -1  d t -2  d t -3  .... n terms ... 
n

f t+1 = forecast for period ‘t + 1’

d t = actual demand for period ‘t’

d t-1 = actual demand for period ‘t - 1’

Simple Average Method is not very popular because it uses

distant past data. The more popular method is the Moving
Average Method
Moving Average Method
This method uses the average of the most recent data
values in the time series as a forecast for the next period.

The term moving indicates that, as the new observations

become available for the time series, it replaces the
oldest observation and the new average is computed. As a
result the average will move as the new observations will
become available.

The first decision we have to take while using moving

average is to decide the number of periods
3 Period Moving Average

Assuming three periods

1
f t 1   dt  d t -1  d t -2 
3

1
Sub t 3, f 4   d 3  d 2  d1 
3

1
Sub t 4, f 5   d4  d3  d2
3
4 Period Moving Average

Assuming four periods

1
f t 1   dt  d t -1  d t -2  d t -3 
4

1
Sub t 4, f 5   d 4  d 3  d 2  d1 
4

1
Sub t 5, f 6   d5  d4  d3  d2 
4
Selecting the number of Periods in Moving Average Method

An important consideration in selecting a forecasting method is the accuracy of

forecast. We define forecast error as ≡ e t = │dt - ft│

We want the forecast errors to be small. The MSE (Mean Squared Error) is an
often used measure of the accuracy of a forecasting method.
1  e2
MSE n t

For a particular time series, different lengths of moving averages, will affect
the accuracy of the forecast. One possible approach in moving average to
choosing the number of periods to be included, is to use trial & error to
identify the number of periods that minimizes MSE.

We must forecast the next value in the time series using the number of data
values that minimizes the MSE for the historical times series.
Weighted Moving Average Method

In moving average method, each observation in the

calculation receives the same weight. One variation
known as Weighted Moving Average, involves selecting
different weights for different data values and then
computing a weighted average of the most recent data.

To use the weighted moving average method, we must

select the number of periods and then choose weights for
each of the periods.

In general, if we believe that the recent past is a better

predictor of the future than the distant past, larger
weights should be given to the more recent periods.
Weighted Moving Average Method
f t 1 α 0 d t  α 1 d t -1  α 2 d t -2  .............

f t+1 = forecast for period ‘t + 1’

α 0 , α1 , α 2 are the weights associated with d t , d t  1 and d t  2

respectively such that:

 αi 1
If we believe what is more recent is more relevant then:
α 0 α1 α 2 ......

The only mandatory requirement in selecting the weights is that their sum
must be equal to 1. The two decisions we have to take while using
weighted moving average is to decide the number of periods & the value
of weights
3 Period Weighted Moving Average

f t  1 α 0 d t  α 1 d t - 1  α 2 d t - 2 (Assuming 3 periods)

Sub t 3, f 4 α 0 d 3  α 1 d 2  α 2 d 1

Sub t 4, f 5 α 0 d 4  α 1 d 3  α 2 d 2

One set of suggested values of  are:

3 2 1
α 0  ; α1  ; α2 
6 6 6

For a particular time series, different values of weight, will affect the
accuracy of the forecast. One possible approach is to use trial & error to
identify the values of weight that minimizes MSE.
 Exponential Smoothing Method
Exponential smoothing is a special case of the weighted moving averages
method in which we select only one weight – the weight for the most recent
observation. The weights of the other data values are automatically computed
using geometric progression and get smaller & smaller as observations move
further into the past. The advantage of exponential smoothing over weighted
moving average is that we have to select only one value of  in exponential
smoothing whereas in weighted average we select more than one value for
example 0, 1 and so on

f t  1 α d t  (1 - α) f t

f t + 1 = forecast for period ‘t + 1’

 = smoothing constant such that (0    1)

dt = actual demand for period ‘t’

ft = forecast for period ‘t’

Exponential Smoothing Method
Initialization of calculation - Assume that f1 = d1 or f2 = d1

f3 onwards can be calculated using the formula as indicated below:

f t 1 α d t  (1 - α ) f t

Substitute t = 2, f3 =  d2 + (1 - ) f2

Substitute t = 3, f4 =  d3 + (1 - ) f3

Although any value of  between 0 & 1 are acceptable, some

values will yield a better forecast than the others.
Selecting the value of  in Exponential Smoothing

The criterion we use to determine the desirable value for the

smoothing constant  is the same as the criterion for determining the
number of periods of data to include in the moving average
calculation. i.e. we choose the value of  that minimizes MSE through
trial & error.

Spreadsheet packages like Excel spreadsheets are an effective aid in

choosing a good value for  for exponential smoothing & selecting
weights for weighted moving average method & selecting number of
periods for moving average method.

With the time series data & forecasting formulas in the spreadsheets,
one can experiment with different values of , or moving average
weights or number of periods & choose the values providing the
smallest MSE.
Using Trend Projection in Forecasting

When we use regression analysis to relate the variable that we

want to forecast to other variables that are supposed to influence
that variable, it becomes a causal forecasting method. Using
regression analysis for trend projection is not a causal forecasting
method because only past values of sales, the variable being
forecast, are used.
If we use simple linear regression to fit to a linear trend to sales
time series then, the sales isn’t actually causally related to time,
instead time is a surrogate for variables to which sales is actually
related, but which are unknown or too difficult or costly to
Using Trend Projection in Forecasting

The type of time series pattern for which the trend

projection method is applicable shows a consistent
increase or decrease over time. It is not stable. So the
smoothing methods are not applicable.

The trend component does not follow each and every up

and down movement. Rather the trend component
reflects the gradual shifting, for example, growth of
time series values.
Trend Analysis in Time Series
Using Regression for Trend Analysis

Ŷt  β̂1  β̂ 2 X t

Ŷt is the estimated Sales value

β̂1 is the intercept term

β̂ 2 is the slope of the regression line

Xt is the time variable (surrogate variable)

β̂1 Y - β̂ 2 X

β̂ 2 
xy ; where x X - X and y Y - Y
x 2
Using Trend and Seasonal Components in
Forecasting

Many situations in business & economics involve period to period

comparisons for example use of electricity consumption is down 3%
from the previous month. Care must be exercised in using such
information because whenever a seasonal influence is present, such
comparisons usually are not meaningful. For example the use of
electricity consumption is down 3% might be a seasonal effect
associated with a decrease in the use of air conditioning & not because
of long-term decline in the use of electrical power. In fact, after
adjusting for seasonal effect, we might even find that the use of electric
power has increased.

Removing the seasonal effect from a time series is known as

deseasonalising the time series. After we do so, period to period
comparisons are more meaningful and can help identify whether a trend
exists.
Using Trend and Seasonal Components in
Forecasting
 Using Trend and Seasonal Components in
Forecasting

The following is a suggested approach:

1. Compute Centered Moving Average

2. Obtain Seasonal Index

3. Obtain deseasonalised sales

4. Use deseasonalised sales data to obtain trend using regression

analysis.

5. Use seasonal index to adjust trend projection.

Qualitative Methods in Forecasting

1. Delphi Method is a Group Consensus of experts physically

separated & unknown to each other. The Questionnaire
process produces a narrow spread of opinion of experts
concurrence.

2. Expert Judgement of a single or group of experts who come

together & brainstorm to provide good forecast for situation
where past conditions are not likely to hold in the future.

3. Sales Force Estimates are based on daily / weekly / monthly

sales estimates collected for each territory & collated for
sales at State / National level

4. Market Research Survey Estimates are based on field survey

using well defined Questionnaire & summary data analysis.