Ensemble Forecasting: A Better Default for Business Forecasting

Why no single model is uniformly best across a portfolio of heterogeneous series, and how combination delivers the accuracy of the best available model without requiring you to identify it.

Three Approaches to Model Selection

Per-Series Selection

Choose the winning model for each series from its own validation window. Across 293 out-of-sample backtests it identified the eventual best model 26% of the time, and combination beat it in 200 of them.

Portfolio-Level Selection

Rank the models on pooled validation evidence and apply the winner across every series. It reaches parity with combination, but only once roughly 150 backtests have accumulated, which is around 50 qualifying series.

Model Combination

Average the member forecasts each period with the extremes trimmed. It matches the best available single model from the first backtest and lands in the bottom three of eleven methods in 1% of backtests against 17.4%.

Open data and code

Check every number in this paper

The data, forecasts, and code behind this research are public. Regenerate every table and figure, then compare your run to ours. It covers 99 daily series from 65 organizations and 293 out-of-sample backtests, free to use under CC BY 4.0.

To hear when we publish the next paper, leave your email below.