Reproduce “Combining Forecasts using Meta-Learning: A Comparative Study for Complex Seasonality” 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
Grzegorz Dudek
Grzegorz Dudek systematically evaluates various machine learning models as meta-learners for combining forecasts of time series with complex seasonality. The study demonstrates enhanced forecasting accuracy over simple averaging, with Random Forest achieving the lowest Mean Absolute Percentage Error (MAPE) of 1.52% for hourly electricity load data across 35 European countries.
Key insights
- Sophisticated meta-learning strategies, particularly stacking with machine learning models, consistently offer substantial accuracy improvements over basic forecast averaging for time series with complex seasonality.
- The choice of meta-learner is crucial, with non-linear models like Random Forest, kNN, and LSTM proving more effective at capturing intricate relationships among base forecasts and the true target.
- For meta-learners, global training often performs as well as or better than local variants, except for specific LSTM configurations (e.g., recent past sequences).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.