Reproduce “Multi-layer Stack Ensembles for Time Series Forecasting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Nathanael Bosch, Oleksandr Shchur, Nick Erickson, Michael Bohlke-Schneider, Caner Türkmen

The authors introduce multi-layer stacking for time-series forecasting, combining forecasts from base models through several stackers and then aggregating those stackers. They benchmark 33 combination methods on 50 datasets. For probabilistic forecasts, its average relative error is 0.945, versus 0.973 for selecting one stacker and 1.000 for median aggregation; results vary by dataset, and training takes longer.

Key insights

  • Ensembling consistently improves predictive accuracy in time series forecasting tasks, and sophisticated, learning-based aggregation methods significantly outperform simple averages, challenging the "forecast combination puzzle."
  • The multi-layer stacking framework effectively combines the strengths of diverse single-layer stackers, leading to robust and superior performance by adaptively leveraging their individual contributions.
  • Optimal performance of multi-layer stacking requires retraining L2 models on all validation data after the L3 aggregator model has been fitted, despite incurring additional computational cost.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.