Session 07

Forecasting Techniques
 Introduction
Forecasting is a very difficult task, both in the
short run and in the long run.
Analysts search for patterns or relationships in
historical data and then make forecasts.
There are two problems with this approach:
It is not always easy to undercover historical patterns or
relationships.
It is often difficult to separate the noise, or random behavior,
from the underlying patterns.
Some forecasts may attribute importance to patterns that are in
fact random variations and are unlikely to repeat themselves.
There are no guarantees that past patterns will continue in
the future.
 Principles of Forecasting
Many types of forecasting models that differ in
complexity and amount of data & way they
generate forecasts:
1. Forecasts are rarely perfect
2. Forecasts are more accurate for grouped
data than for individual items
3. Forecast are more accurate for shorter than
longer time periods
 Types of Forecasting Methods
Decide what needs to be forecast
Level of detail, units of analysis & time horizon required
Evaluate and analyze appropriate data
Identify needed data & whether its available
Select and test the forecasting model
Cost, ease of use & accuracy
Generate the forecast
Monitor forecast accuracy over time
 Forecasting Methods:
An Overview
There are many forecasting methods available,
and there is little agreement as to the best
forecasting method.
The methods can be divided into three groups:
1. Judgmental methods
2. Extrapolation (or time series) methods
3. Econometric (or causal) methods
The first method is basically nonquantitative; the
last two are quantitative.
 Extrapolation Models
Extrapolation models are quantitative models that use past data of
a time series variable to forecast future values of the variable.
Many extrapolation models are available:
Trend-based regression
Autoregression
Moving averages
Exponential smoothing
All of these methods look for patterns in the historical series and
then extrapolate these patterns into the future.
Complex models are not always better than simpler models.
Simpler models track only the most basic underlying patterns and can
be more flexible and accurate in forecasting the future.
 Quantitative Methods
Time Series Models
Assumes information needed to generate a forecast is contained
in a time series of data
Assumes the future will follow same patterns as the past
Causal Models or Associative Models
Explores cause-and-effect relationships
Uses leading indicators to predict the future
Housing starts and appliance sales
 Time Series Models
Forecaster looks for data patterns as
Data = historic pattern + random variation
Historic pattern to be forecasted:
Level (long-term average) data fluctuates around a constant
mean
Trend data exhibits an increasing or decreasing pattern
Seasonality any pattern that regularly repeats itself and is of a
constant length
Cycle patterns created by economic fluctuations
Random Variation cannot be predicted
 Time Series Models

Naive: Ft +1 = At
The forecast is equal to the actual value observed during
the last period good for level patterns
Simple Mean: Ft +1 = A t / n
The average of all available data - good for level patterns
Moving Average:
Ft +1 = A t / n
The average value over a set time period
(e.g.: the last four weeks)
Each new forecast drops the oldest data point & adds a
new observation
More responsive to a trend but still lags behind actual
data
 Time Series Models cont

Weighted Moving Average: Ft +1 = C t A t

All weights must add to 100% or 1.00

e.g. Ct .5, Ct-1 .3, Ct-2 .2 (weights add to 1.0)

Allows emphasizing one period over others; above

indicates more weight on recent data (Ct=.5)

Differs from the simple moving average that weighs all

periods equally - more responsive to trends
 Time Series Models cont

Exponential Smoothing: Ft +1 = A t + (1 )Ft

Most frequently used time series method because of ease
of use and minimal amount of data needed
Need just three pieces of

data to start:
Last periods forecast (Ft)
Last periods actual value (At)

Select value of smoothing coefficient, ,between 0 and 1.0
If no last period forecast is available, average the last few
periods or use naive method

Higher values (e.g. .7 or .8) may place too much weight
on last periods random variation
 Time Series Problem

Determine forecast for periods 7 & 8 Period Actual

2-period moving average 1 300
4-period moving average 2 315
2-period weighted moving average with t-1 3 290
weighted 0.6 and t-2 weighted 0.4
4 345
Exponential smoothing with alpha=0.2 and
the period 6 forecast being 375 5 320
6 360
7 375
8
 Time Series Problem Solution

Period Actual 2-Period 4-Period 2-Per.Wgted. Expon. Smooth.

1 300

2 315

3 290

4 345

5 320

6 360

7 375 340.0 328.8 344.0 372.0

8 367.5 350.0 369.0 372.6

 Forecasting trend problem: a company uses exponential smoothing with trend to forecast
usage of its lawn care products. At the end of July the company wishes to forecast sales
for August. July demand was 62. The trend through June has been 15 additional gallons of
product sold per month. Average sales have been 57 gallons per month. The company
uses alpha+0.2 and beta +0.10. Forecast for August.

Smooth the level of the series:

S July = A t + (1 )(S t 1 + Tt 1 ) = (0.2 )(62 ) + (0.8)(57 + 15 ) = 70

Smooth the trend:

TJuly = (St St 1 ) + (1 )Tt 1 = (0.1)(70 57 ) + (0.9 )(15) = 14.8

Forecast including trend:

FITAugust = S t + Tt = 70 + 14.8 = 84.8 gallons
 Causal Models

Often, leading indicators can help to predict changes in

future demand e.g. housing starts
Causal models establish a cause-and-effect relationship
between independent and dependent variables
A common tool of causal modeling is linear regression:
Additional related variables may require multiple
regression modeling

Y = a + bx
 Linear Regression

Identify dependent (y) and

independent (x) variables
b=
XY ((X ) Y ) Solve for the slope of the line
X 2 ((X ) X )

b=
XY n XY

2
X nX
2

Solve for the y intercept

a = Y bX
Develop your equation for the
trend line
Y=a + bX
 Correlation Coefficient
How Good is the Fit?
Correlation coefficient (r) measures the direction and strength of the
linear relationship between two variables. The closer the r value is to
1.0 the better the regression line fits the data points.

n( XY ) ( X )( Y )
r=
( )
n X ( X ) * n Y (Y ) ( )
2 2
2 2

4(28,202 ) 189(589 )
r= = .982
4(9253) - (189) * 4(87,165 ) (589 )
2 2

r 2 = (.982 ) = .964
2

r2

Coefficientr 2 of determination ( ) measures the amount of variation in

the dependent variable about its mean that is explained by the
regression line. Values of ( ) close to 1.0 are desirable.
Time Series Patterns

18
 Linear Trend Line

A time series technique that computes a forecast

with trend by drawing a straight line through a
set of data using this formula:
Y = a + bx where
Y = forecast for period X
X = the number of time periods from X = 0
A = value of y at X = 0 (Y intercept)
B = slope of the line
 Forecasting Trend

Basic forecasting models for trends compensate for the lagging that
would otherwise occur
One model, trend-adjusted exponential smoothing uses a three step
process
Step 1 - Smoothing the level of the series

S t = A t + (1 )(S t 1 + Tt 1 )
Step 2 Smoothing the trend

Tt = (S t S t 1 ) + (1 )Tt 1
Forecast including the trend
FITt +1 = S t + Tt
 Forecasting Seasonality

Calculate the average demand per season

E.g.: average quarterly demand
Calculate a seasonal index for each season of each
year:
Divide the actual demand of each season by the
average demand per season for that year
Average the indexes by season
E.g.: take the average of all Spring indexes, then of all
Summer indexes, ...
 Seasonality cont
Forecast demand for the next year & divide by the
number of seasons
Use regular forecasting method & divide by four for
average quarterly demand
Multiply next years average seasonal demand by
each average seasonal index
Result is a forecast of demand for each season of next year
 Combining Forecasts
This method combines two or more forecasts to obtain
the final forecast.
The reasoning is simple: The forecast errors from
different forecasting methods might cancel one another.
Forecasts that are combined can be of the same general
type, or of different types.
The number of forecasts to combine and the weights to
use in combining them have been the subject of several
research studies.
 Components of Time Series Data
If observations increase or decrease regularly through time, the time
series has a trend.
Linear trendoccurs if the observations increase by the same amount from
period to period.
Exponential trendoccurs when observations increase at a tremendous rate.
S-shape trendoccurs when it takes a while for observations to start
increasing, but then a rapid increase occurs, before finally tapering off to a
fairly constant level.
 Components of Time Series Data
If a time series has a seasonal component, it exhibits
seasonalitythat is, the same seasonal pattern tends to
repeat itself every year.
 Components of Time Series Data

A time series has a cyclic component when business cycles affect

the variables in similar ways.
The cyclic component is more difficult to predict than the seasonal
component, because seasonal variation is much more regular.
The length of the business cycle varies, sometimes substantially.
The length of a seasonal cycle is generally one year, while the length
of a business cycle is generally longer than one year and its actual
length is difficult to predict.
 Components of Time Series Data

Random variation (or noise) is the unpredictable component that

gives most time series graphs their irregular, zigzag appearance.
A time series can be determined only to a certain extent by its trend,
seasonal, and cyclic components; other factors determine the rest.
These other factors combine to create a certain amount of
unpredictability in almost all time series.
 Testing for Randomness
All forecasting models have the general form shown in the
equation below:

The fitted value is the part calculated from past data and any
other available information.
The residual is the forecast error.
The fitted value should include all components of the original
series that can possibly be forecast, and the leftover residuals
should be unpredictable noise.
The simplest way to determine whether a time series of
residuals is random noise is to examine time series graphs
of residuals visuallyalthough this is not always reliable.
 Testing for Randomness
Some common nonrandom patterns are shown below.
 The Runs Test
The runs test is a quantitative method of testing
for randomness. It is a formal test of the null
hypothesis of randomness.
First, choose a base value, which could be the average
value of the series, the median value, or even some
other value.
Then a run is defined as a consecutive series of
observations that remain on one side of this base
level.
If there are too many or too few runs in the series, the null
hypothesis of randomness can be rejected.
 Autocorrelation
Another way to check for randomness of a time series of residuals
is to examine the autocorrelations of the residuals.
An autocorrelation is a type of correlation used to measure whether
values of a time series are related to their own past values.
In positive autocorrelation, large observations tend to follow large
observations, and small observations tend to follow small observations.
The autocorrelation of lag k is essentially the correlation between the
original series and the lag k version of the series.
Lags are previous observations, removed by a certain number of periods from
the present time.
To lag a time series in a spreadsheet by one month, push down the series
by one row, as shown below.
 Regression-Based Trend Models
Many time series follow a long-term trend
except for random variation.
This trend can be upward or downward.
A straightforward way to model this trend is to
estimate a regression equation for Yt, using time t
as the single explanatory variable.
The two most frequently used trend models are
the linear trend and the exponential trend.
 Linear Trend
A linear trend means that the time series variable
changes by a constant amount each time period.
The equation for the linear trend model is:
The interpretation of b is that it represents the expected
change in the series from one period to the next.
If b is positive, the trend is upward.
If b is negative, the trend is downward.
The intercept term a is less important: It literally
represents the expected value of the series at time t = 0.
A graph of the time series indicates whether a linear
trend is likely to provide a good fit.
 Exponential Trend
An exponential trend is appropriate when the time series changes
by a constant percentage (as opposed to a constant dollar amount)
each period.
The appropriate regression equation contains a multiplicative error
term ut:
This equation is not useful for estimation; for that, a linear equation
is required.
You can achieve linearity by taking natural logarithms of both sides of
the equation, as shown below, where a = ln(c) and et = ln(ut).

The coefficient b (expressed as a percentage) is approximately the percentage

change per period. For example, if b = 0.05, then the series is increasing by
approximately 5% per period.
If a time series exhibits an exponential trend, then a plot of its
logarithm should be approximately linear.
 Moving Averages Forecasts
One of the simplest and the most frequently used
extrapolation models is the moving averages model.
A moving average is the average of the observations in the past
few periods, where the number of terms in the average is the
span.
If the span is large, extreme values have relatively little effect on
the forecasts, and the resulting series of forecasts will be much
smoother than the original series.
For this reason, this method is called a smoothing method.
If the span is small, extreme observations have a larger effect on
the forecasts, and the forecast series will be much less smooth.
Using a span requires some judgment:
If you believe the ups and downs in the series are random noise, use a
relatively large span.
If you believe each up and down is predictable, use a smaller span.
 Exponential Smoothing Forecasts
Exponential smoothing bases its forecasts on a
weighted average of past observations, with
more weight on the more recent observations.
There are many variations of exponential
smoothing, including:
Simple exponential smoothingappropriate for a
series with no pronounced trend or seasonality
Holts methodappropriate for a series with trend
but no seasonality
Winters methodappropriate for a series with
seasonality (and possibly trend)
 Simple Exponential Smoothing
Every exponential model has at least one smoothing constant,
which is always a number between 0 and 1.
Simple exponential smoothing has a single smoothing constant
denoted by .
The level of the series at time t (Lt) is an estimate of where the
series would be at time t if there were no random noise.
The simple exponential method is defined by the following two
equations:

The second equation says that the k-period-ahead forecast, Ft+k, made
of Yt+k in period t is essentially the most recently estimated level, Lt.
 Measuring Forecast Error

Forecasts are never perfect

Need to know how much we should rely on
our chosen forecasting method
Measuring forecast error:
E t = A t Ft
Note that over-forecasts = negative errors
and under-forecasts = positive errors
 Measuring Forecasting Accuracy

Mean Absolute Deviation (MAD) MAD =

actual forecast
measures the total error in a forecast n
without regard to sign
Cumulative Forecast Error (CFE) CFE = (actual forecast )
Measures any bias in the forecast

(actual - forecast )2

Mean Square Error (MSE) MSE =

Penalizes larger errors n
CFE
Tracking Signal TS =
MAD
Measures if your model is working
 Accuracy & Tracking Signal Problem: A company is comparing the accuracy of
two forecasting methods. Forecasts using both methods are shown below along
with the actual values for January through May. The company also uses a tracking
signal with 4 limits to decide when a forecast should be reviewed. Which
forecasting method is best?

Method A Method B
Month Actual Fcast Error Cum. Tracking Fcast Error Cum. Tracking
sales Signal Error Signal
Error

Jan. 30 28 2 2 2 27 2 2 1

Feb. 26 25 1 3 3 25 1 3 1.5
March 32 32 0 3 3 29 3 6 3
April 29 30 -1 2 2 27 2 8 4
May 31 30 1 3 3 29 2 10 5

MAD 1 2
MSE 1.4 4.4
 Forecasting Software
Spreadsheets
Microsoft Excel, Quattro Pro, Lotus 1-2-3
Limited statistical analysis of forecast data
Statistical packages
SPSS, SAS, NCSS, Minitab
Forecasting plus statistical and graphics
Specialty forecasting packages
Forecast Master, Forecast Pro, Autobox, SCA