Simulation of Stochastic Rolling Horizon Forecast Behavior with Applied Outlier Correction to Increase Forecast Accuracy

Customer provided demand forecasts are often applied for production planning in order to create production orders for the respective materials. However, practical observations show that based on the customers’ planning system, there might be temporary fluctuations in the demand forecast values that have a high magnitude which lead to instabilities in production plans with negative impact on production system performance. In the forecast data this behavior is measurable in form of numerical outliers.

For planning the production steps the demand information is important, but when forecasts are not accurate and rely on outliers, they introduce nervousness into production system with negative impact on production system performance.

To reduce this forecast introduced uncertainty within the production system researchers developed two simple outlier correction methods and extended the forecast generation process.

In this paper the focus lies on the investigation of the performance and applicability of the proposed outlier corrections.

Demand Forecast Model and Outlier Correction

To increase forecast accuracy, researchers tried to detect outliers in the forecast data and correct them to smoothen the demand forecasts. They used 2 correction methods:

  • Method 1, when outliers are identified based on the mean and variance of the final order amounts and the average order amount is used instead of the original forecast if an outlier is detected;
  • Method 2, when outliers are identified similar to method 1, but the last forecast value for the respective due date (in M1 the average order amount is used) is used to replace the original forecast.

To evaluate the performance of the developed outlier correction model, researchers used AnyLogic’s capabilities of supply chain forecasting software and created an discrete event simulation model. In detail, only at specific periodic points in time, a change in the forecast is triggered in the model. The model has implemented the forecast generation process described in Zeiml et al. (2020) and is extended with the already described outlier generation as well as the outlier correction methods.


The results show that both outlier correction methods perform well and provide the expected smoothening of the forecasts values which consequently can reduce nervousness in the production system. However, results show that method 2 significantly performance better than method 1. Furthermore, the results show that outlier correction might add an additional source of uncertainty if no outliers occur and therefore, high threshold values for outlier identification are advantageous.

In future research the outlier identification could be extended to be based on the mean and variance of the forecast updates with respect to the periods before delivery instead of the final order amounts.

Related posts