Distributions of forecasting errors of forecast combinations: implications for inventory management

Distributions of forecasting errors of forecast combinations: implications for inventory management

Devon Barrow, Nikolaos Kourentzes

 

Key Points

1. Forecasting errors do not respect well normal distribution in practice, even though combination forecasting does better than base methods.

2. Combination forecasts have a better correlation between the in-sample error variance and out-of-sample variance, therefore more 'reliable' when deciding Safety Stock.

3. Combination forecasts have a better accuracy and bias compared to base methods.

4. Among combination forecasting methods, Median is better than Mean, in contrast to the practice.

5. The conclusion is based on the experiment done by the authors. It's not guaranteed to be 'correct'.

 

Abstract

The study examines the forecast error distributions of base and combination forecasts and their implications for inventory performance.

 

Introduction and background

1. Many inventory control systems assume that forecast errors are normally distributed and unbiased. This is often violated in practice.

2. It's well recognized that combining forecasts improves forecasting accuracy. However, no much result exists on the impact of combining forecasts on the error distribution, which directly influence the choice of Safty Stock (based on the standard deviation of forecasting error instead of the std of sales).

3. It has been observed that the difference between the properties of in- and out-of-sample errors is quite large and variable, which implies that the SS that we set may seem pretty good with respect to the samples but may result in stock out in pratice.

4. Some common forecast combination methods (Simple mean/median, Optimal, regression... ) are introduced, most of which can be theoretically represented as a linear combination of forecasts.

Results

1. On forecasting accuracy and bias: Theta and MAPA perform best. The latter employs multiple temporal aggregated series of the original time series, to achieve better estimation of the various time series components. Among combination methods, the simple Median method is the best.

2. On normality: even though combinations methods better respect the normality assumption, the normality is violated very frequently. (less than 50% of cases.) Tested using Shapiro-Wilk test, which tests the deviation from normal distribution.

3. On the correlation of in-sample variance of error and out-sample variance, combination methods perform better than base methods. The means for combination methods, in-sample variance is an appropriate estimation of the out-of-sample variance.

4. When considering the impact on Safety Stock, when not assuming normal distribution on samples but choose an empirical distribution by model fitting, combination methods are more beneficial on SS.

Conclusion

1. Combination methods improve forecasting accuracy and bias and also result in more normally distributed errors.

2. Empirical error distribution of the combination forecasts is a good approximation of the out-of-sample one, which is not obvious on base methods. It's also better than assuming a normal distribution.

3. A sidary observation is that the widely used MEAN method did not perform well (compared to MEDIAN) in the presence of irregular data. This is relevant to inv management where promotions are common.