Calibrated parameters
Window & seasonality sensitivity
Same series, different calibration windows and seasonal handling — checks whether the OU fit and the OU-vs-GBM backtest are stable or whether they shift once history is shorter, per AZR-511/AZR-504's calibration-window question.
| Window | Months | κ | θ (c/kWh) | σ | Half-life | Sanity | OU RMSE | GBM RMSE | Backtest winner | Margin | Review flag |
|---|
Is OU always the best model? No — not assumed here. The table above backtests OU against GBM on true out-of-sample one-step-ahead error for this series, at multiple window lengths. Validated against synthetic ground-truth series (see
test_calibration.py in the repo): one-step RMSE often can't cleanly separate OU from GBM on 24–60 months of monthly data, especially for fast-mean-reverting series (half-life near the 1-month sampling interval) or short 2yr windows. Naively auto-switching the Monte Carlo engine to whichever model wins the backtest produced explosive, unstable forecast tails on a genuinely mean-reverting series — so this page keeps OU as the default central case (per the platform spec: OU = DCF central case, GBM = stress-test only) and treats a GBM backtest win as a review flag, not an automatic switch.