Portfolio of Predictions

Forecasts are assets. Forecast errors are "returns." Combination weights are portfolio weights. This demo builds the mean–squared-error-optimal combination of several toy forecasters using real mean–variance portfolio math — matrix inversion on the actual mean vector and covariance matrix of their errors — and lets it go short a forecaster when that's the error-minimizing move. Based on arXiv:2308.15384 — "Hedging Forecast Combinations With an Application to the Random Forest".

— Computing…

Simple Average

—

MSE (equal weights)

Best Long-Only Mix

—

MSE (weights ≥ 0)

Hedged Optimal

—

MSE (shorting allowed)

Combination Weights by Method

Mean Squared Error — Lower is Better

Monte-Carlo Verification — Running RMSE of Combined Forecast (converges to the theory)

Simple average (simulated) Best long-only (simulated) Hedged optimal (simulated) dashed = closed-form theoretical RMSE

Underlying Error Statistics (what the optimizer actually sees)

Mean forecast error (bias) vector μ

Covariance matrix Σ of forecast errors