You are on page 1of 23

ECON 332

Business Forecasting Methods


Chapter 8
Combining Forecast Results
Prof. Kirti K. Katkar
8-2
Combining Forecast Results
Why combine forecast results rather than using a single
method?
• 83% of expert forecasters believe that combining forecast
results from different models improves forecast accuracy
• Optimal individual forecasting model offers lowest RMSE and
statistically “best” forecasts; combing forecasts offers the
forecast improvement over the individual methods
• Combining forecast results tend to be inclusionary
– They include all the variables from models that would be discarded
otherwise to choose statistically best forecast
– They include all the relationships from models that would be discarded
otherwise to choose statistically best forecast
• Major requirement for combining forecast results is that
forecasts to be combined need to be unbiased
– They cannot all consistently overestimate or underestimate the true
value
8-3

Combined Forecast

To ensure unbiased individual forecasts


•Choose models based on different methods
•Have different forecasters develop different models / forecasts
8-4

Techniques for Choosing Weights for


Combined Forecasts
• Bates and Granger Method
– Major assumption is that individual forecasts are consistent over time
and are determined through minimization of variance of forecast errors
over time
– Technique is used in combining two individual forecasts
– If weight k is assigned to first forecast, the second forecast is assigned
the weight of (1-k), with k being defined as

k = (σ22 –ρσ1σ2) / (σ12 - σ22 - 2 ρσ1σ2)

where k – weight assigned to the first forecasting model


σ12 - Variance of errors in the first forecasting model
σ22 - Variance of errors in the second forecasting model
ρ – the coefficient of correlation between errors in the first
model and those in the second
8-5

Techniques for Choosing Weights for Combined


Forecasts (Contd.)
• Adaptive weighting method
– Allows weights to adapt or change from period to period
– Implied assumption is that a forecasting model may not have a constant
performance over time
– Technique is used in combining two individual forecasts
– The adaptive set of weights are defined as
α1,T+1 = Σ(e2t2/ (e1t2 +e2t2)), t = T-v, T-v+1,…., T
where α1,T+1 = The weight assigned to forecast model 1 in period T=1
eit = The error made by forecast model I for period t
v = The choice variable = # of periods you want to include in the
adaptive weighting procedure, typically 4 for quarterly data
T = Total # of periods for which there is a history of forecast error
– Of course, the weight assigned for forecast model 2 would be
α2,T+1 = 1 - α1,T+1
8-6
Charles Nelson Regression Weighting Technique
• To minimize error of combined forecast, an optimal linear composite
forecast would be:
F* = b1F(1) + b2F(2), for each corresponding value of time t
where F* = Optimal combined forecast
F(1) = Forecast from model 1
F(2) = Forecast from model 2
bi = Weight assigned to ith forecast i=1,2
• The weights bi are determined by a two-step regression process and the
sum of bi roughly equals 1
8. Perform a standard multiple regression using actual values as a
dependent variable and forecasted values from methods 1 and 2 as
independent variables, say
A = a + b1F(1) + b2F(2)
2. The value of intercept a should be zero if there is no bias in the
combined forecast. A two tailed t test is used to ensure that a = 0
11. Assuming a = 0, redo regression by forcing it through the origin. This
will yield the optimal weights and forecast.
F* = b1F(1) + b2F(2)
8-7

Bessler and Brandt Empirical Evidence

• Forecasts from individual models are not likely to be the most


accurate forecasts viz-a-vie weighted combined forecasts
• Even with no prior forecast performance (accuracy tracking),
it may make sense to combine forecasts at least using the equal
weight method
• If prior forecasting accuracy track is available, the user should
weight forecasts based on prior performance i.e. place higher
weight on methods that yield more accurate forecasts
8-8
RMSEs for Gap Sales Forecasts
8-9

Figure 8-2
8-10

Figure 8-3
8-11

Table 8-1
8-12

Table 8-2
8-13

Table 8-3
8-14

Table 8-4
8-15

Table 8-4 (continued)


8-16

Table 8-5
8-17

Table 8-5 (continued)


8-18

Table 8-6
8-19

Table 8-6 (continued)


8-20

Table 8-7
8-21

Table 8-8
8-22

Table 8-8 (continued)


8-23

Table 8-8 (continued)

You might also like