EUROPEAN POWER MARKET RESEARCH

German wind forecast validation

XGBoost residual calibration | 11 expanding walk-forward folds | paper wind-to-price strategy

View source
No reliable incremental edge. Forecast skill is within noise and residual strategy P&L is statistically weak. The large public-ramp result diagnoses proxy structure; it is not evidence of free executable profit.

Forecast MAE skill vs SMARD

-- Loading validation result

Residual net P&L / trade

-- Loading strategy result

Public-ramp proxy / trade

-- Loading benchmark result

Positive residual folds

-- Loading fold concentration

FORECAST VALIDATION

Fold-by-fold MAE against the public forecast

Loading fold results...

Read: a model with repeatable edge should beat the public forecast across regimes. Here, October is the clearest failure and the pooled gain is indistinguishable from noise.

STRATEGY EVIDENCE

Cumulative proxy P&L on one shared scale

Same validation hours, 0.5 EUR/MWh cost, one MW hourly clips.

The scale is the finding: the public rule reaches 137,188.39 EUR while the residual signal finishes at 188.72 EUR. Because both use DA(t) minus DA(t-24), that gap shows the proxy rewards public-information repricing; it does not establish an unarbitraged strategy.

Residual P&L by validation fold

THRESHOLD ROBUSTNESS

Net P&L per trade across the pre-defined grid

The selected 1,500 MW threshold is marked; every point uses the same walk-forward folds.

No stable residual threshold: the rule peaks at 1,000 MW and turns negative above 2,000 MW. The best grid point is an unadjusted five-choice maximum, not holdout evidence.

AI ANALYST REVIEW

Evidence-bounded downstream interpretation

Loading the saved review...

Loading analyst conclusion

Forecast assessment

    Strategy assessment

      Risk flags

        Invalidation conditions

          Production next steps

            Boundary: the LLM summarizes deterministic outputs after the run. It cannot alter QA, forecasts, trades, or metrics, and every displayed evidence value comes from the original May snapshot.

            EXECUTION CAVEAT

            Research proxy, not an executable backtest

            P&L uses DA(t) minus DA(t-24), not auction-to-intraday or imbalance settlement marks. Weather rows and 24-hour lags are also not pinned to strict D-1 noon gate closure. Production validation requires issue-time features and executable settlement prices.