Forecast v6 — Four-Leg Ensemble
A four-leg ensemble that blends v1 (v11 P50), v4 (EPEX-CZ), a global CatBoost GBM and a per-slot LEAR (LASSO) model. Each delivery day uses a walk-forward, accuracy-weighted blend — every leg's weight is its inverse-MAE over strictly prior days, so better recent models get more say. The two GBM/linear legs are retrained weekly and stored; v1/v4 are their frozen forecasts. Scored on the same cleared OTE 15-min grid, with v5 shown as the baseline.
Mean over 0 scored days
Production v1 × v4 ensemble, same days
No overlap yet
Avg weights —/—/—/— (v1/v4/cat/lear)
Loading runs…
Blend: for each delivery day D, v6 = Σ wₖ·legₖ at every 15-min step over the four legs, where wₖ = (1/MAEₖ) / Σ(1/MAEⱼ) computed over the trailing scored days strictly before D (≥3 days required, else a 25/25/25/25 cold-start). This is the inverse-MAE recipe that won the offline four-leg study (≈3% MAE improvement over v5), applied walk-forward so it never sees the day it predicts.
Legs & leakage:v1 is the calibrated v11 median; v4 is its frozen EPEX-CZ point forecast. CatBoost (one global GBM over the v11 feature frame + slot) and LEAR (one LASSO per 15-min slot) are retrained weekly on a trailing 180-day window — cutoff at the week's first Prague-midnight, so they only ever train on data knowable before the predicted week. Actuals are the cleared OTE 15-min price, so scores fill in as the DAM clears.