Operations Management

Session 11-13 – Forecasting

Dr. Ramesh Krishnan

IIM Kozhikode
3-1
[email protected]
Features Common to All Forecasts
1. Techniques assume some underlying causal system that
existed in the past will persist into the future
2. Forecasts are not perfect
3. Forecasts for groups of items are more accurate than
those for individual items
4. Forecast accuracy decreases as the forecasting horizon
increases

3-6
Steps in the Forecasting Process
1. Determine the purpose of the forecast
2. Establish a time horizon
3. Obtain, clean, and analyze appropriate data
4. Select a forecasting technique
5. Make the forecast
6. Monitor the forecast errors

3-9
Forecasting Approaches
 Qualitative forecasting
 Qualitative techniques permit the inclusion of soft information
such as:
 Human factors
 Personal opinions
 Hunches
 These factors are difficult, or impossible, to quantify
 Quantitative forecasting
 These techniques rely on hard data
 Quantitative techniques involve either the projection of historical
data or the development of associative methods that attempt to use
causal variables to make a forecast

3-10
Qualitative Forecasts
 Forecasts that use subjective inputs such as opinions from consumer
surveys, sales staff, managers, executives, and experts
 Executive opinions
 A small group of upper-level managers may meet and collectively develop a
forecast
 Salesforce opinions
 Members of the sales or customer service staff can be good sources of
information due to their direct contact with customers and may be aware of
plans customers may be considering for the future
 Consumer surveys
 Since consumers ultimately determine demand, it makes sense to solicit input
from them
 Consumer surveys typically represent a sample of consumer opinions
 Other approaches
 Managers may solicit 0pinions from other managers or staff people or outside
experts to help with developing a forecast.
 The Delphi method is an iterative process intended to achieve a consensus

3-11
Time-Series Forecasts
 Forecasts that project patterns identified in recent
time-series observations
 Time-series – a time-ordered sequence of observations
taken at regular time intervals
 Assume that future values of the time-series can be
estimated from past values of the time-series

3-12
Time-Series Behaviors
 Trend
 Seasonality
 Cycles
 Irregular variations
 Random variation

3-13
Time-Series Behaviors
 Trend
 A long-term upward or downward movement in data
 Population shifts
 Changing income

 Seasonality
 Short-term, fairly regular variations related to the calendar or time of day
 Restaurants, service call centers, and theatres all experience seasonal demand.
 Cycle
 Wavelike variations lasting more than one year
 These are often related to a variety of economic, political, or even agricultural conditions.

 Random Variation
 Residual variation that remains after all other behaviors have been accounted for

Random Variation

3-14
M T W T F
Time-Series Forecasting - Naïve Forecast
 Naïve forecast
 Uses a single previous value of a time series as the basis
for a forecast
 The forecast for a time period is equal to the previous
time period’s value
 Can be used with
 A stable time series
 Seasonal variations – same as last season
 Trend

3-15
Time-Series Forecasting - Averaging
 These techniques work best when a series tends to vary
about an average
 Averaging techniques smooth variations in the data
 They can handle step changes or gradual changes in the
level of a series
 Techniques
1. Moving average
2. Weighted moving average
3. Exponential smoothing

3-16
Moving Average
 Technique that averages a number of the most recent
actual values in generating a forecast
n

A t −i
At − n + ... + At − 2 + At −1
Ft = MA n = i =1
=
n n
where
Ft = Forecast for time period t
MA n = n period moving average
At −i = Actual value in period t − i
n = Number of periods in the moving average

 As new data become available, the forecast is updated

by adding the newest value and dropping the oldest
and then re-computing the average 3-17
Example – Moving Average

3-18
Weighted Moving Average
 The most recent values in a time series are given more
weight in computing a forecast
 The choice of weights, w, is somewhat arbitrary and
involves some trial and error.
Ft = wt ( At ) + wt −1 ( At −1 ) + ...+ wt − n ( At − n )
where
wt = weight for period t , wt −1 = weight for period t − 1, etc.
At = the actual value for period t , At −1 = the actual value for period t − 1, etc.

(Training and test set/ cross validation

3-19
3-20
Exponential Smoothing
 A weighted averaging method that is based on the
previous forecast plus a percentage of the forecast
error
Ft = Ft −1 +  ( At −1 − Ft −1 )
where
Ft = Forecast for period t
Ft −1 = Forecast for the previous period
 = Smoothing constant
At −1 = Actual demand or sales from the previous period

3-21
Example – Exponential Smoothing

3-22
Linear Trend

3-23
Correlation Coefficient
 Measures the relative strength of the linear
relationship between two variables
 Unit-less
 Ranges between –1 and 1
 The closer to –1, the stronger the negative linear
relationship
 The closer to 1, the stronger the positive linear
relationship
 The closer to 0, the weaker any positive linear
relationship

3-24
 Scatter Plots of Data with Various Correlation
Coefficients
Y Y Y

X X X
r = -1 r = -.6 r=0
Y
Y Y

X X X
r = +1 r = +.3 r=0 3-25
Linear Correlation
Linear relationships Curvilinear relationships

Y Y

X X

Y Y

X X3-26
Linear Correlation
Strong relationships Weak relationships

Y Y

X X

Y Y

X X3-27
Linear Correlation
No relationship

X
3-28
Correlation Coefficient
 Formula

n( xy ) − ( x )( y )
r=
( )
n  x 2 − ( x )
2
( )
n  y 2 − ( y )
2

 r2, square of the correlation coefficient

 A measure of the percentage of variability in the values of y that is
“explained” by the independent variable
 Ranges between 0 and 1.00

3-29
 Linear Trend
 A simple data plot can reveal the existence and nature
of a trend
 Linear trend equation
Ft = a + bt n  ty −  t  y
b=
n t − ( t )
2
where 2

Ft = Forecast for period t

a=
 y − b t or y − bt
a = Value of Ft at t = 0 n
where
b = Slope of the line n = Number of periods
t = Specified number of time periods from t = 0 y = Value of the time series

3-30
3-31
Trend-Adjusted Exponential Smoothing
 The trend adjusted forecast consists of two
components
 Smoothed error
 Trend factor
TAFt +1 = S t + Tt
where
S t = Previous forecast plus smoothed error
Tt = Current trend estimate
S t = TAFt +  ( At − TAFt )
Tt = Tt −1 +  (TAFt − TAFt −1 − Tt −1 )
• Alpha and beta are smoothing constants
• Trend-adjusted exponential smoothing has the ability to respond to
changes in the trend
For initialization, we can consider the previous period demand as forecast and 0 as trend or perform linear
trend and identify or average change per period in first four period for trend and 4th period demand as 3-32
forecast (expo component)
3-34
Associative Forecasting Techniques
 Associative techniques are based on the
development of an equation that summarizes the
effects of predictor variables
 Predictor variables – variables that can be used to
predict values of the variable of interest
 Home values may be related to such factors as home and
property size, location, number of bedrooms, and number of
bathrooms

3-35
Simple Linear Regression
 Regression – a technique for fitting a line to a set of
data points
 Simple linear regression – the simplest form of
regression that involves a linear relationship between
two variables
 The object of simple linear regression is to obtain an equation
of a straight line that minimizes the sum of squared vertical
deviations from the line (i.e., the least squares criterion)

3-36
Least Squares Line
yc = a + bx
where
yc = Predicted (dependent) variable
x = Predictor (independent) variable
b = Slope of the line
a = Value of yc when x = 0 (i.e., the height of the line at the y intercept)
and
n( xy ) − ( x )( y )
b=
( )
n  x 2 − ( x )
2

a=
 y − b x
or y − b x
n
where
n = Number of paired observations
3-38
Simple Linear Regression Assumptions
1. Variations around the
line are random
2. Deviations around the
average value (the line)
should be normally
distributed
3. Predictions are made
only within the range of
observed values

This is the reason why we generally consider prediction interval instead of a fixed number
3-39
(Second hand car price estimator) https://orangebookvalue.com/
Conditions for the SRM ? Checklist
 Is the association between y and x linear?
 Have lurking variables been ruled out?
 Are the errors evidently independent?
 Are the variances of the residuals similar?
 Are the residuals nearly normal?

3-40
Assumptions-Summary

Assumption Diagnostic Problem

Linearity Scatter Plots Nonlinear

Independence Plots showing trend Autocorrelation

Constant Variance Residual Plots Heteroscedasticity

Normality Quantile Plots Outliers, Skewness

3-41
Example:
Customized manufacturing is a growth industry as businesses
adjust to worldwide competition

Key to success in this industry is the ability to provide rapid turn-around. When a client calls
with an order for 225 special metal brackets, managers need to be able to respond quickly
with an estimate of when the brackets can be delivered and how much the order will cost.
That’s a job well-suited to regression
Reference: Stine and Foster; Chapter 21 and 22 Class 2: Simple Regression Model

The explanatory variable gives the sizes

of 47 orders for similar brackets, and
the response is the time (in minutes)
needed for fabrication.

Estimated Production Time

= 172 + 2.44 Number of Units
Conditions for the SRM-Linearity

Reference: Stine and Foster; Chapter 21 and 22 Class 2: Simple Regression Model
Conditions for the SRM- Residual Plot

Reference: Stine and Foster; Chapter 21 and 22 Class 2: Simple Regression Model

Linearity condition is satisfied; no pattern in the residuals.

Conditions for the SRM- Similar Variance

Similar variances condition is satisfied. Check the plot of

residuals versus x for any fan shaped pattern (none
visible).
Conditions for the SRM- Normality

Reference: Stine and Foster; Chapter 21 and 22 Class 2: Simple Regression Model

Nearly normal condition satisfied.

Procedure

Plot y versus x and verify the association appears linear.

If pattern is linear, fit the least squares regression line and
obtain residuals
Plot residuals against explanatory variable- plot should
have no pattern. Note presence of outliers as well.
Reference: Stine and Foster; Chapter 21 and 22 Class 2: Simple Regression Model

If data measured over time, check for timeplot for

dependence
Inspect histogram and quantile plot of residuals to check
normality.
Issues to Consider:
 Always plot the line to verify that a linear relationship
is appropriate
 The data may be time-dependent
 If they are
 use analysis of time series
 use time as an independent variable in a multiple regression
analysis
 A small correlation may indicate that other variables
are important

3-48
Forecast Accuracy and Control
 Allowances should be made for forecast errors
 It is important to provide an indication of the extent to
which the forecast might deviate from the value of the
variable that actually occurs
 Forecast errors should be monitored
 Error = Actual – Forecast
 If errors fall beyond acceptable bounds, corrective
action may be necessary

3-51
Forecast Accuracy Metrics

MAD =
 Actual − Forecast
t t MAD weights all errors
n evenly

2
σ Actualt − Forecast t MSE weights errors according
MSE =
𝑛 to their squared values

Actualt − Forecast t
 Actualt
100
MAPE weights errors
MAPE = according to relative error
n

3-52
When to Use MAE:
• When you want a straightforward measure of accuracy that is easy to
interpret.
• When outliers are present, but you do not want them to have a large
influence on the overall error metric.
• MAE gives an average error magnitude in the same units as the output
variable, making it intuitively easier to interpret than MSE, which is in
squared units.

When to Use MSE:

•When you want to emphasize larger errors more than smaller ones since
squaring the errors before averaging them penalizes larger errors.
•When the data does not contain many outliers that can skew the squared error.
Forecast Error Calculation
Actual Forecast (A-F)
Period
(A) (F) Error |Error| Error2 [|Error|/Actual]x100

1 107 110 -3 3 9 2.80%

2 125 121 4 4 16 3.20%

3 115 112 3 3 9 2.61%

4 118 120 -2 2 4 1.69%

5 108 109 1 1 1 0.93%

Sum 13 39 11.23%

n=5 n=5 n=5

MAD MSE MAPE

= 2.6 = 7.8 = 2.25%

3-54
Monitoring the Forecast
 Tracking forecast errors and analyzing them can provide useful
insight into whether forecasts are performing satisfactorily
 Sources of forecast errors:
 The model may be inadequate due to
a. omission of an important variable
b. a change or shift in the variable the model cannot handle
c. the appearance of a new variable
 Random variations may have occurred
 Control charts are useful for identifying the presence of non-
random error in forecasts
 Tracking signals can be used to detect forecast bias

3-55
Tracking Signal

3-56
Choosing a Forecasting Technique
 Factors to consider
 Cost
 Accuracy
 Availability of historical data
 Availability of forecasting software
 Time needed to gather and analyze data and prepare a
forecast
 Forecast horizon

3-58