EURO 2025 Leeds
Abstract Submission

74. An Ensemble-Based Machine Learning Approach to Increase the Robustness of Forecast Combination Weights

Invited abstract in session MD-34: Advancements of OR-analytics in statistics, machine learning and data science 1, stream Advancements of OR-analytics in statistics, machine learning and data science.

Monday, 14:30-16:00
Room: Michael Sadler LG10

Authors (first author is the speaker)

1. Veronika Wachslander
Chair of Business Administration and Business Informatics, Catholic University of Eichstaett-Ingolstadt
2. Thomas Setzer
Chair of BA and Business Informatics, Catholic University of Eichstätt-Ingolstadt

Abstract

Forecasting economic or business figures is often the basis for corporate planning, and combining individual forecasts by computing their weighted average regularly increases predictive accuracy. A huge amount of literature deals with the determination of beneficial combination weights and one popular approach is the estimation of so-called optimal weights (OW), minimizing the mean squared error on available past data. However, this and other approaches that learn weights from data usually perform worse on unseen data than simply assigning equal weights (EW). The reason can be seen in the typically limited amount of data available for the estimation, so that weights are (over)fit on random structures that often do not occur on new data, in turn leading to strongly misfit weights and high errors.
One widely used approach to reduce overfitting is to estimate OW and shrink these towards EW. This technique requires the tuning of a shrinkage hyperparameter, which is subject to uncertainty. Unfortunately, this uncertainty usually also causes higher errors than EW on unseen data in empirical settings.
We introduce a machine learning method using subsampling to learn and shrink combination weights. The method aims at reducing overfitting, while still considering generalizable structures in training data by learning more stable weights that should better fit new, unseen data. Experiments with synthetic data show promising results compared to common weighting and tuning strategies.

Keywords

Status: accepted

Back to the list of papers