You are on page 1of 31

The Art of Forecasting

MITSloan

Costs of Forecasting

IBM continues to struggle with shortages in the Think Pad


line. WSJ 5/2/94
IBM sells out new Aptiva PC. Shortage may cost millions
WSJ 10/7/94
Dell stock plunges. Dellwas sharply off in its forecast
of demand WSJ 8/93
Liz Claiborne said earnings decline is consequence of
[unanticipated] excess inventories. WSJ 7/15/93
Toyota believes it can save $100M/3a [with] accurate
ordering and inventory management. Its unbelievable
the number of (after market) parts that are thrown away
each year. Automotive News 2/26/01
MITSloan

Store Mark-Downs
% of Dollar Sales

26%
22%
18%
14%
10%
6%
1970

1975

1980

1985

1990

Source: Fisher, Marshall L. et al. Making Supply Meet Demand in an Uncertain World. Boston, MA:
Harvard Business Review, 1994. Reprint No. 94302. and National Retail Federation

MITSloan

Forecasting Basics

Forecasts are usually wrong

Forecasts should contain error measure.

Aggregate Forecasts are more accurate


The longer the horizon, the lower the
accuracy
Common Sense Compatibility

Include other known Information


MITSloan

Basic Taxonomy

Qualitative
(subjective)

Expertise Based
Delphi Method
The Sage
Sales Force
Customer Surveys
Long Term
Complex Scenarios

Quantitative
(objective)

Causal Models

Y=f(X1, X2, , Xn)

Time Series Models

Yt =f(Yt-1, Yt-2,, Yt-n)

MITSloan

Causal Models

Use Data other than the Series


predicted
Principal Tool: Regression Analysis

Y = a0 + a1X1 + a2X2+ anXn


Estimates ai to minimize Sum of Least
Squares
Y

MITSloan

Time Series Models

Prediction based exclusively on


previously observed Values
General Idea: Detect Patterns!
Short Term Demand Prediction
Prevalent Tool In Operations

MITSloan

Time Series Patterns

Random

Trend

Linear (default) or Nonlinear

Seasonality

No Pattern

Repetition at Fixed Intervals

Cyclic

Long Term Economy


MITSloan

The Basic Concept

Ft = a n Dt n
n =1

Ft : Forecast for Period t , made in Period t-1


Di : Observed Demand in Period i < t
a i : Weights associated with Period i

Question: How to assign Weights?


MITSloan

Stationary Series
Moving Averages
&
Exponential Smoothing
MITSloan

Moving Averages
Dt 1 + Dt 1 + ... + Dt N
1 N
Ft = Dt n =
N
N n =1
a i = N1 for i < N , a i = 0 elsewhere

Simple Average of Last N Periods


No Pattern
Large N Stability up
Small N Responsiveness up
MITSloan

Exponential Smoothing
Ft = a Dt 1 + (1 a ) Ft 1
= Ft 1 a (Ft 1 Dt 1 )

= Ft 1 a et 1

New Forecast = Weighted Average of

et 1 The last Error

Last Demand and Last Forecast

Small a Stability up
Large a Responsiveness up
MITSloan

Moving Average vs.


Exponential Smoothing

Commonalities

Stationary Process
1 Parameter (a or N)
Lag behind Trend

Differences

ES: All past data


MA: Last N periods

ES: Less data

Outliers washed out

Last Forecast *
Demand

MA: Last N data

MITSloan

Trends
Exponential Smoothing with
Linear Trend
Regression Analysis

MITSloan

Exponential Smoothing with


Linear Trend
Ft +1 = St + Gt

St = a Dt + (1 a ) [St 1 + Gt 1 ]

= Ft

Gt = b (St St 1 ) + (1 b ) Gt 1
Ft + = St + Gt

St The Intercept or Status Quo


Gt The Trend
MITSloan

Regression Analysis
Sum of Least Squares
Min Sum of

Y
a0

Y = a0 + a1*X

X
MITSloan

Regression in Excel

Formal Regression Analysis


1.
2.
3.
4.
5.

Tools, Data Analysis, Regression


Input Y Range: Dependent Variable
Input X Range: Independent Variable
Specify where you want Output
Output is Table with Regression Statistics

MITSloan

Trend Tool

=Trend(y-range,x-range,x-value)
Output: Trend Estimate for x-value
Y-range: Observed Data (Demand)
X-range: Independent Variable (Time)
X-value: Value (Date) for which to
estimate Y (Demand)

MITSloan

Non Linear Trend

Regression

Same Procedure as Linear Regression


Difference:
Change Time to
2
Time
ln(Time)
etc..

MITSloan

Seasonal Patterns

Basic Idea:

Assign Weights cn to the N Periods


c n = N
Adjust Forecast by Weights

Caution:

A Season is the time until a Pattern repeats

I.e. a week is a Season with 7 Periods


A year is a Season with 4 Periods
MITSloan

Error Measures

Mean Absolute Deviation MAD


MAD =

1
n

F D
t

i =1

Mean Square Deviation MSD


MSD =

Bias

1
n

F D
t

i =1

1
n

(F D )
i =1

MITSloan

How to choose the right


forecasting technique?
Chambers, Mullick & Smith

MITSloan

Demand

Demand and Product Life


Cycles

Product
Development

Steady
State

Introduction
Growth

Decline

Time

Demand Predictions are dependent on


Life Cycle
MITSloan

Decisions Made During PLC

Product Development

Development Effort
Market Entry
Product Specs

Product Introduction

facility size
supply chain design

Analysis

delphi / expert
Comparisons
QFD
Analysis

market tests
consumer survey
life cycle analysis
MITSloan

Decision Made During PLC

Growth

capacity expansion
statistical tech
production planning
promotions

Steady State

production planning
inventory models

Analysis

Causal Models
Simulation

Analysis

time series
causal models
MITSloan

Next Week

Monday

Voluntary Lead-Offs
Download Inventory Problems from
SloanSpace in Handouts folder

MITSloan

Sale of SKU 367890,


a seasonal product
PREDICTED

Sales

OBSERVED

ACTUAL
Feb

Apr

Jun

Aug

Oct

Dec
MITSloan

Demand

Random Patterns

Time
MITSloan

Demand

Increasing Linear Trend

Time
MITSloan

Demand

Curve Linear Trend

Time
MITSloan

Demand

Seasonal & Linear Trend

Time
MITSloan

You might also like