2223. Shifting Forecast Combination Weights Estimated on Subsamples towards Full Sample Weight Determination
Invited abstract in session WB-6: Predictive Analytics: Forecast Combination & Hyperparameter Optimization, stream Analytics, Data Science, and Forecasting.
Wednesday, 10:45-12:15Room: H9
Authors (first author is the speaker)
| 1. | Veronika Wachslander
|
| Chair of Business Administration and Business Informatics, Catholic University of Eichstaett-Ingolstadt |
Abstract
Combining forecasts of different predictive models or forecasters has proven to be powerful for increasing predictive accuracy. Numerous techniques exist to determine the combination weights assigned to the models or forecasters. One common approach estimates optimal weights (OW) by minimizing the mean squared error (MSE) on a provided set of past (training) data. However, the estimated weights are influenced by random structures and tend to overfit limited training data. An approach for reducing this overfitting, while still incorporating differences in predictive ability is to shrink OW towards equal weights (EW). However, this technique typically suffers from structural distortions (i.e., strong overfitting) of the initially learned OW, which cannot be remedied by shrinking unless full shrinkage to EW is applied.
Modern machine learning-based ensemble methods, which estimate parameters (weights in this context) on multiple training subsamples and aggregate the results, are less prone to overfitting compared to estimations on the full training sample.
We introduce a method that aims at learning more robust combination weights by estimating OW on multiple subsamples of training data and shrinking these weights towards EW to obtain an ensemble of decorrelated weight vectors. Subsequently, these vectors are linearly shifted to a controllable degree towards (less overfitted) weights learned on the full training sample. The final combination weights are obtained by averaging the ensemble of shifted weight vectors.
The method is evaluated and compared to OW and EW on synthetic datasets with varying training and subsample sizes, different shrinkage and shifting levels, and different variance-covariance structures of forecasters.
Keywords
- Forecasting algorithms
- Machine Learning
- Robustness
Status: accepted
Back to the list of papers