When Forecast Accuracy Fails: Rank Correlation and Decision Quality in Multi-Market Battery Storage Optimization

Battle Count: 67

Highly relevant to quantitative trading in energy markets. The paper directly addresses the core question of how forecast quality translates into trading performance, demonstrating that rank correlation (Kendall τ) rather than pointwise error (MAE) is the primary predictor of intraday dispatch value. The tau-sufficiency threshold provides a concrete, actionable criterion for forecast model selection in trading systems. The capacity allocation analysis reveals that in multi-market systems, the allocation decision dominates forecast quality as a revenue driver. The adaptive bidding framework (HMM+SAC) and regime detection are directly applicable to trading strategy development. The findings challenge the standard practice of evaluating trading forecasts by MAE/RMSE, with implications for any ordinal decision problem in trading.

High complexity due to the hierarchical three-layer architecture requiring: (1) stochastic linear programming for weekly capacity allocation with PICASSO compliance constraints, (2) rolling MPC LP for daily scheduling with exponentially declining forecast weights, (3) DP Rolling Intrinsic for 15-minute intraday dispatch, (4) HMM regime detection with Baum-Welch training and Viterbi decoding, (5) SAC reinforcement learning agent for adaptive bid selection, (6) LightGBM autoregressive forecasting with sequential rollout, (7) synthetic forecast generation with binary search calibration for tau-sufficiency experiments, (8) walk-forward out-of-sample evaluation across 6 years of 15-minute data. The system requires careful handling of SoC constraints, market gate closures, and cross-layer physical consistency. However, the modular architecture allows component replacement.

All data sources are publicly available (Regelleistung.net, SMARD, Swissgrid, ENTSO-E, Open-Meteo ERA5, ICE/yfinance). Methodology is described in detail with explicit equations. Walk-forward out-of-sample evaluation is specified. Battery parameters are fully stated. However, no code repository is linked, and some implementation details (e.g., exact LGBM hyperparameters, DP state discretization) are not fully specified. The synthetic forecast generation procedure is described but the binary search calibration details could be more explicit.