Module 4.

Forecasting
MGS3100

Forecasting
Forecasting
Quantitative

Qualitative

Causal Model

Expert Judgment
Trend

Delphi Method

Time series

Grassroots
Stationary
Trend

Trend + Seasonality

Quantitative Forecasting
--Forecasting based on data and models
Casual Models:
Price
Population
Advertising

Causal
Model

Year 2000
Sales

Time Series Models:

Sales1999
Sales1998
Sales1997

Time Series
Model

Year 2000
Sales

Causal forecasting
Regression
Find a straight line that fits the data best.
Best line!

Intercept

y = Intercept + slope * x (= b0 + b1x)

Slope = change in y / change in x

Causal Forecasting Models

Curve Fitting: Simple Linear Regression
One Independent Variable (X) is used to predict one
Dependent Variable (Y): Y = a + b X
Given n observations (Xi, Yi), we can fit a line to the
overall pattern of these data points. The Least
Squares Method in statistics can give us the best a
and b in the sense of minimizing (Yi - a - bXi)2:

XiYi

Xi Yi

2
X
i

( Xi ) 2

Regression formula is an optional learning objective

5

 Curve Fitting: Simple Linear Regression

Find the regression line with Excel
Use Function:
a = INTERCEPT(Y range; X range)
b = SLOPE(Y range; X range)
Use Solver
Use Excels Tools | Data Analysis | Regression

Curve Fitting: Multiple Regression

Two or more independent variables are used to
predict the dependent variable:
Y = b0 + b1X1 + b2X2 + + bpXp
Use Excels Tools | Data Analysis | Regression

Time Series Forecasting Process

Look at the data
(Scatter Plot)
Observations from the
scatter Plot

Forecast using one or

more techniques

Evaluate the technique

and pick the best one.

Techniques to try

Ways to evaluate

Data is reasonably
stationary
(no trend or seasonality)

Heuristics - Averaging methods

Naive
Moving Averages
Simple Exponential Smoothing

MAD
MAPE
Standard Error
BIAS

Data shows a consistent

trend

Regression
Linear
Non-linear Regressions (not
covered in this course)

MAD
MAPE
Standard Error
BIAS
R-Squared

Data shows both a trend and

a seasonal pattern

Classical decomposition
Find Seasonal Index
Use regression analyses to find
the trend component

MAD
MAPE
Standard Error
BIAS
R-Squared
7

Evaluation of Forecasting Model

BIAS - The arithmetic mean of the errors
BIAS

(Actual - Forecast) Error

n

n is the number of forecast errors

Excel: =AVERAGE(error range)

Mean Absolute Deviation - MAD

| Actual - Forecast | | Error |

MAD

No direct Excel function to calculate MAD

Evaluation of Forecasting Model

Mean Square Error - MSE
(Actual - Forecast) (Error)
MSE
2

Excel: =SUMSQ(error range)/COUNT(error range)

Standard error is square root of MSE

Mean Absolute Percentage Error - MAPE

| Actual - Forecast |
*100%

Actual
MAPE
n

R2 - only for curve fitting model such as regression

In general, the lower the error measure (BIAS, MAD,
MSE) or the higher the R2, the better the forecasting
model
9

Stationary data forecasting

Nave
I sold 10 units yesterday, so I think I will sell 10 units
today.
n-period moving average
For the past n days, I sold 12 units on average.
Therefore, I think I will sell 12 units today.
Exponential smoothing
I predicted to sell 10 units at the beginning of yesterday;
At the end of yesterday, I found out I sold in fact 8 units.
So, I will adjust the forecast of 10 (yesterdays forecast)
by adding adjusted error ( * error). This will compensate
over (under) forecast of yesterday.
10

Nave Model
The simplest time series forecasting
model
Idea: what happened last time (last year,
last month, yesterday) will happen again
this time
Nave Model:

Algebraic: Ft = Yt-1

Yt-1 : actual value in period t-1

Ft : forecast for period t

Spreadsheet: B3: = A2; Copy down

11

Moving Average Model

Simple n-Period Moving Average

Sum of actual values in previous n periods

n
Y
Y
Y
t

1
t

2
tn
=
n

F
t

Issues of MA Model
Nave model is a special case of MA with n = 1
Idea is to reduce random variation or smooth data
All previous n observations are treated equally (equal
weights)
Suitable for relatively stable time series with no trend or
seasonal pattern

12

Smoothing Effect of MA Model

Longer-period moving averages (larger n) react to

actual changes more slowly
13

Moving Average Model

Weighted n-Period Moving Average
F =w Y
w Y
w Y
t
1 t 1
2 t2
n tn

Typically weights are decreasing:

w1>w2>>wn
Sum of the weights = wi = 1

Flexible weights reflect relative importance of

each previous observation in forecasting
Optimal weights can be found via Solver

14

Weighted MA: An Illustration

Month Weight Data
August 17% 130
September 33% 110
October 50% 90
November forecast:
FNov = (0.50)(90)+(0.33)(110)+(0.17)(130)
= 103.4

15

Exponential Smoothing
Concept is simple!
Make a forecast, any forecast
Compare it to the actual
Next forecast is
Previous forecast plus an adjustment
Adjustment is fraction of previous forecast error

Essentially
Not really forecast as a function of time
Instead, forecast as a function of previous actual and
forecasted value

16

Simple Exponential Smoothing

A special type of weighted moving average
Include all past observations
Use a unique set of weights that weight recent observations
much more heavily than very old observations:

0 1

weight
Decreasing weights
given
to older
observations

(1 )
(1 ) 2
(1 ) 3

17
Toda
y

Simple ES: The Model

Ft Yt 1 (1 )Yt 2 (1 ) 2 Yt 3
Ft Yt 1 (1 )Yt 2 (1 a )Yt 3

Ft Yt 1 (1 ) Ft 1
New forecast = weighted sum of last period
actual value and last period
forecast
: Smoothing constant
Ft :Forecast for period t
Ft-1:

Last period forecast

Yt-1:

Last period actual value

18

Simple Exponential Smoothing

Properties of Simple Exponential
Smoothing
Widely used and successful model
Requires very little data
Larger , more responsive forecast; Smaller
, smoother forecast (See Table 13.2)
best can be found by Solver
Suitable for relatively stable time series
19

Time Series Components

Trend
persistent upward or downward pattern in a time series

Seasonal
Variation dependent on the time of year
Each year shows same pattern

Cyclical
up & down movement repeating over long time frame
Each year does not show same pattern

Noise or random fluctuations

follow no specific pattern
short duration and non-repeating

20

Time Series Components

Cycle

Trend

Random
movement
Time

Seasonal
pattern

Time

Demand

Time

Trend with
seasonal pattern

21

Time

Trend Model
Curve fitting method used for time series
data (also called time series regression
model)
Useful when the time series has a clear trend
Can not capture seasonal patterns
Linear Trend Model: Yt = a + bt
t is time index for each period, t = 1, 2, 3,
7
6
5
4
3
2
1
0
1

10

22

Pattern-based forecasting - Trend

Regression Recall Independent Variable X, which is now
time variable e.g., days, months, quarters, years etc.

Find a straight line that fits the data best.

Best line!

Intercept

y = Intercept + slope * x (= b0 + b1x)

Slope = change in y / change in x

23

Pattern-based forecasting Seasonal

Once data turn out to be seasonal,
deseasonalize the data.
The methods we have learned (Heuristic methods and
Regression) is not suitable for data that has
pronounced fluctuations.

Make forecast based on the deseasonalized data

Reseasonalize the forecast
Good forecast should mimic reality. Therefore, it is
needed to give seasonality back.

24

Pattern-based forecasting Seasonal

Example (SI + Regression)
Actual data

Deseasonalized data

Deseasonalize

Forecast

Reseasonalize

25

Pattern-based forecasting Seasonal

Deseasonalization
Deseasonalized data = Actual / SI
Reseasonalization
Reseasonalized forecast
= deseasonalized forecast * SI

26

Seasonal Index
Whats an index?
Ratio
SI = ratio between actual and average demand

Suppose
SI for quarter demand is 1.20
Whats that mean?
Use it to forecast demand for next fall
So, where did the 1.20 come from?!

27

Calculating Seasonal Indices

Quick and dirty method of calculating SI
For each year, calculate average demand
Divide each demand by its yearly average
This creates a ratio and hence a raw index
For each quarter, there will be as many raw indices
as there are years
Average the raw indices for each of the quarters
The result will be four values, one SI per quarter

28

Classical decomposition
Start by calculating seasonal indices
Then, deseasonalize the demand
Divide actual demand values by their SI values
y = y / SI
Results in transformed data (new time series)
Seasonal effect removed

Forecast
Regression if deseasonalized data is trendy
Heuristics methods if deseasonalized data is stationary

Reseasonalize with SI
29

Causal or Time series?

What are the difference?
Which one to use?

30

Can you
describe general forecasting process?
compare and contrast trend, seasonality and
cyclicality?
describe the forecasting method when data is
stationary?
describe the forecasting method when data
shows trend?
describe the forecasting method when data
shows seasonality?
31