Forecasting in Supply Chain

The Role of Forecasting

Forecasting is a vital function and affects every significant management decision.

◦ Finance and accounting use forecasts as the basis for budgeting and cost control.
◦ Marketing relies on forecasts to make key decisions such as new product planning and
personnel compensation.
◦ Production uses forecasts to select suppliers; determine capacity requirements; and
drive decisions about purchasing, staffing, and inventory.

Different roles require different forecasting approaches.

◦ Decisions about overall directions require strategic forecasts.
◦ Tactical forecasts are used to guide day-to-day decisions.
 Forecasting and Decoupling Point
Decoupling point: Point at which inventory is stored, which allows SC to
operate independently.

The choice of the decoupling point in a SC is strategic.

Forecasting helps determine the level of inventory needed at the decoupling

points.

The decision will be affected by the error produced in the forecast and the type
of product (easily inventoried or easily perishable).
 Types of Forecasting
There are four basic types of forecasts.
◦ Qualitative
◦ Time series analysis (primary focus of this chapter)
◦ Causal relationships
◦ Simulation

Time series analysis is based on the idea that data relating to past demand can
be used to predict future demand.
 Seven Steps in Forecasting

1. Determine the use of the forecast

2. Select the items to be forecasted
3. Determine the time horizon of the forecast
4. Select the forecasting model(s)
5. Gather the data needed to make the forecast
6. Make the forecast
7. Validate and implement results
 Forecasting Principles

 There is no such things as 100% reliable forecast

 Forecast for more accurate for the larger group of items
 Forecast are more accurate over shorter time periods
 Every forecast should be accompanied by an estimate of forecast error for the planning purposes
 Hierarchy of Forecasting

Top Down Approach Bottom Up Approach

Forecast Management Process
 Qualitative Forecasting Techniques
Generally used to take advantage of expert knowledge.

Useful when judgment is required, when products are new, or if the firm has little experience in a
new market.

Examples
◦ Market research
◦ Panel consensus
◦ Historical analogy
◦ Delphi method
 Model Selection

Choosing an appropriate forecasting model depends upon

◦ Time horizon to be forecast

◦ Data availability
◦ Accuracy required
◦ Size of forecasting budget
◦ Availability of qualified personnel
 Time-Series Components

Trend Cyclical

Seasonal Random
 Components of Demand
Trend
component

Demand for product or service

Seasonal peaks

Actual demand
line

Average demand
over 4 years

Random variation
| | | |
1 2 3 4
Time (years)
 Seasonal Component

► Regular pattern of up and down fluctuations

► Due to weather, customs, etc.
► Occurs within a single year
 Cyclical Component

► Repeating up and down movements

► Affected by business cycle, political, and economic factors
► Multiple years duration
► Often causal or
associative
relationships

0 5 10 15 20
 Random Component

► Erratic, unsystematic, ‘residual’ fluctuations

► Due to random variation or unforeseen events
► Short duration
and nonrepeating

M T W T
F
 Forecasting Method Selection Guide

Forecast
Forecasting Method Amount of Historical Data Data Pattern
Horizon

6 to 12 months; weekly Stationary (i.e., no

Simple moving average Short
data are often used trend or seasonality)

Weighted moving average

5 to 10 observations
and simple exponential Stationary Short
needed to start
smoothing
5 to 10 observations
Exponential smoothing
needed to start Stationary and trend Short
with trend

Stationary, trend, and Short to

Linear regression 10 to 20 observations
seasonality medium
 Simple Moving Average
Forecast is the average of a fixed number of past periods.

Useful when demand is not growing or declining rapidly and no seasonality is present.

Removes some of the random fluctuation from the data.

Selecting the period length is important.

◦ Longer periods provide more smoothing.
◦ Shorter periods react to trends more quickly.
Simple Moving Average Formula
Simple Moving Average – Example

18-19
 Weighted Moving Average
The simple moving average formula implies equal weighting for all periods.

A weighted moving average allows unequal weighting of prior time periods.

◦ The sum of the weights must be equal to one.
◦ Often, more recent periods are given higher weights than periods farther in
the past.

𝐹𝑡 = 𝑤1𝐴𝑡 − 1 + 𝑤2𝐴𝑡 − 2 + …+
𝑤𝑛𝐴𝑡 −𝑛
 Selecting Weights
 Experience and/or trial-and-error are the simplest approaches.

 The recent past is often the best indicator of the future, so weights are generally higher for more
recent data.

 If the data are seasonal, weights should reflect this appropriately.

 Exponential Smoothing
 More recent results weighted more heavily

 The most used of all forecasting techniques

 An integral part of computerized forecasting

Well accepted for six reasons

◦ Formulating an exponential model is relatively easy
◦ The user can understand how the model works
◦ Little computation is required to use the model
◦ Computer storage requirements are small
◦ Tests for accuracy are easy to compute
 Exponential Smoothing

► Form of weighted moving average

► Weights decline exponentially
► Most recent data weighted most
► Requires smoothing constant (α)
► Ranges from 0 to 1
► Subjectively chosen
► Involves little record keeping of past data
 Exponential Smoothing

New forecast = Last period’s forecast + a (Last period’s actual demand – Last period’s forecast)

Ft = Ft – 1 + a(At – 1 - Ft – 1)

where Ft = new forecast

Ft – 1 = previous period’s forecast
a = smoothing (or weighting) constant (0 ≤ a ≤ 1)
At – 1 = previous period’s actual demand
 Exponential Smoothing Example
Predicted demand = 142 Ford Mustangs
Actual demand = 153
Smoothing constant a = .20

New forecast = 142 + .2(153 – 142)

= 144.2 ≈ 144 cars
 Impact of Different 
225 –

Actual  = .5
demand
200 –
Demand

175 –

 = .1
150 – | | | | | | | | |
1 2 3 4 5 6 7 8 9
Quarter
 Impact of Different 
225 –

Actual  = .5
► Chose
200 – high of 
values
demand
when underlying average
Demand

is likely to change
► Choose low values of 
175 –

when underlying average  = .1

is stable|
150 – | | | | | | | |
1 2 3 4 5 6 7 8 9
Quarter
 Choosing 

The objective is to obtain the most accurate forecast no matter the

technique

We generally do this by selecting the model that gives us the lowest

forecast error

Forecast error = Actual demand – Forecast value

= At – Ft
Exponential Smoothing Model

18-29
 Exponential Smoothing Example

Week Demand Forecast

1 820 820
2 775 820
3 680 811
4 655 785
5 750 759
6 802 757
7 798 766
8 689 772
9 775 756
10 760
Exponential Smoothing – Effect of Trends

The presence of a trend in the data causes the exponential smoothing forecast to
always lag behind the actual data

This can be corrected by adding a trend adjustment

◦ The trend smoothing constant is delta (δ)
Example – Exponential Smoothing with Trend Adjustment

Calculate the new forecast, assuming the following:

◦ The previous forecast including trend (FITt-1) is 110 and the previous estimate of the
trend (Tt-1) is 10
◦ α = 0.2 and δ = 0.3
◦ Actual demand for period t-1 is 115

Ft = FITt-1 + α(At-1 – FITt-1) = 110 + 0.2(115-110) = 111.0

Tt = Tt-1 + δ(Ft – FITt-1) = 10 + 0.3(111-110) = 10.3

FITt = Ft + Tt = 111.0 + 10.3 = 121.3

 Choosing Alpha and Delta

Relatively small values for α and δ are common

◦ Usually in the range 0.1 to 0.3

α depends upon how much random variation is present

δ depends upon how steady the trend is

Measurement of forecast error can be used to select values of α and δ to minimize overall forecast
error
 Linear Regression Analysis

 Regression is used to identify the functional relationship between two or more correlated
variables, usually from observed data.
 One variable (the dependent variable) is predicted for given values of the other variable
(the independent variable).
 Linear regression is a special case that assumes the relationship between the variables can
be explained with a straight line.

Y = a + bt
 Least Squares Method
The least squares method determines the parameters a and b such that the sum of the
squared errors is minimized – “least squares”

Quarter Sales Quarter Sales

1 600 7 2,600
2 1,550 8 2,900
3 1,500 9 3,800
4 1,500 10 4,500
5 2,400 11 4,000
6 3,100 12 4,900
 Calculations

1 600 600 1 360,000 801.3

2 1,550 3,100 4 2,402,500 1,160.9

3 1,500 4,500 9 2,250,000 1,520.5

4 1,500 6,000 16 2,250,000 1,880.1

5 2,400 12,000 25 5,760,000 2,239.7

6 3,100 18,600 36 9,610,000 2,599.4

7 2,600 18,200 49 6,760,000 2,959.0

The forecast is extended to periods 13-16 8 2,900 23,200 64 8,410,000 3,318.6

9 3,800 34,200 81 14,440,000 3,678.2

10 4,500 45,000 100 20,250,000 4,037.8

11 4,000 44,000 121 16,000,000 4,397.4

12 4,900 58,800 144 24,010,000 4,757.1

Sum 78 33,350 268,200 650 112,502,500
 Time Series Decomposition
 Chronologically ordered data are referred to as a time series.
 A time series may contain one or many elements.
 Identifying these elements and separating the time series data into these components is known as
decomposition.

The additive model is useful when the seasonal variation is relatively constant over time.
The multiplicative model is useful when the seasonal variation increases over time.
 Seasonal Variation
Seasonal variation may be either additive or multiplicative (shown here with a changing trend).

• The additive model is useful when the seasonal variation is relatively constant over time.
• The multiplicative model is useful when the seasonal variation increases over time.
 Additive or multiplicative
(shown here with a changing trend).
 Determining Seasonal Factors :
Simple Proportions
The seasonal factor (or index) is the ratio of the amount sold during each season divided by the
average for all seasons.

Average Sales for

Season Past Sales Seasonal Factor
Each Season

1000 200
Spring 200 = 250 = 0.8
4 250
1000 350
Summer 350 = 250 = 1.4
4 250
1000 300
Fall 300 = 250 = 1.2
4 250
1000 150
Winter 150 = 250 = 0.6
4 250
Total 1000
 Example

Expected Average Next

Demand Sales for Seasonal Year’s
for Each Season Factor Seasonal
Next Year (1,100y4) Forecast
Spring 275 X 0.8 = 220
Summer 275 X 1.4 = 385
Fall 275 X 1.2 = 330
Winter 275 X 0.6 = 165
1100
Decomposition Using Least Squares Regression
Decompose the time series into its components
◦ Find seasonal component
◦ Deseasonalize the demand
◦ Find trend component

Forecast future values of each component

◦ Project trend component into the future
◦ Multiply trend component by seasonal component
 Decomposition – Steps 1 and 2

Using the data for periods 1-12, apply time series analysis (decomposition, linear
regression, trend estimate & seasonal indices) to forecast for periods 13-16
 Decomposition – Steps 3 and 4
Develop a least squares regression line for the deseasonalized data.
Project the regression line through the period of the forecast.

Regression Results:
Y = 555.0 + 342.2t
Forecast for
periods 13-16
 Decompostion – Step 5
Create the final forecast by adjusting the regression line by the seasonal factor.

Seasonal Forecast (F x
Period Quarter Y from Regression
Factor Seasonal Factor
13 I 5,003.5 0.82 4,102.87
14 II 5,345.7 1.10 5,880.27
15 III 5,687.9 0.97 5,517.26
16 IV 6,030.1 1.12 6,753.71
 Forecast Errors

Forecast error is the difference between the forecast value and what actually
occurred.

All forecasts contain some level of error.

Sources of error
◦ Bias – when a consistent mistake is made
◦ Random – errors that are not explained by the model being used

Measures of error
◦ Mean absolute deviation (MAD)
◦ Mean absolute percent error (MAPE)
◦ Tracking signal
 Forecast Error Measurements
Ideally, MAD will be zero (no • MAPE scales the forecast error to
forecasting error). the magnitude of demand.
Larger values of MAD indicate a less
accurate model.

• Tracking signal indicates whether

forecast errors are accumulating
over time (either positive or
negative errors).
Computing Forecast Error
 Causal Relationship Forecasting
Causal relationship forecasting uses independent variables other than time to predict future
demand.
◦ This independent variable must be a leading indicator.

Many apparently causal relationships are actually just correlated events – care must be taken
when selecting causal variables.
 Multiple Regression Techniques
Often, more than one independent variable may be a valid predictor of future
demand.

In this case, the forecast analyst may utilize multiple regression.

◦ Analogous to linear regression analysis, but with multiple independent variables.
◦ Multiple regression supported by statistical software packages.