Evolving Multilevel Forecast Combination Models - An Experimental Study

This paper provides a description and experimental comparison of different forecast combination techniques for the application of Revenue Management forecasting for Airlines. In order to benefit from the advantages of forecasts predicting seasonal demand using different forecast models on different aggregation levels and to reduce the risks of high noise terms on low level predictions and overgeneralization on higher levels, various approaches based on combination of many predictions are presented and experimentally compared. We propose to evolve combination structures dynamically using Evolutionary Computing approaches. The evolved structures are not only able to generate predictions representing well balanced and stable fusions of methods and levels, they are also characterised by high adaptive capabilities. The focus on different levels or methods of forecasting may change as well as the complexity of the combination structure depending on changes in parts of the input data space in different data aggregation levels. Significant forecast improvements have been obtained when using the proposed dynamic multilevel structures.