Orderbook Feature Learning and Asymmetric Generalization in Intraday Electricity Markets

Battle Count: 86

Highly relevant for quantitative trading in electricity markets. The paper provides actionable insights for market participants operating across multiple European markets, demonstrating that models trained on liquid markets (Germany) can be deployed in less liquid markets (Austria) with minimal performance loss, while the reverse is catastrophic. The feature selection methodology identifies which orderbook features (price percentiles, extreme prices) are most predictive, directly informing trading signal construction. The probabilistic forecasting framework is essential for risk-aware trading decisions. The asymmetric generalization finding has direct implications for multi-market trading strategy design.

Moderate complexity. The feature extraction pipeline (384 features from orderbook) is well-defined but requires access to raw orderbook data. Feature selection via LASSO/LQR is standard. Model training involves multiple architectures (LQR, QKNN, QLGBM, QXGB, QMLP, QKAN) with Bayesian hyperparameter optimization (Optuna, 100 trials). The generalization assessment requires careful cross-domain experimental design. QKAN implementation is relatively novel and complex. Overall, the pipeline is reproducible with standard ML libraries but requires significant computational resources for full benchmarking.

The paper provides a project page (https://runyao-yu.github.io/AsymGen/), detailed hyperparameter search ranges (Table II), explicit feature definitions (Table I), and clear data splitting protocol (train: 2022-01-01 to 2024-01-01, validation: 2024-01-01 to 2024-07-01, test: 2024-07-01 to 2025-01-01). Uses Optuna for hyperparameter tuning with 100 trials. However, the actual code repository availability and orderbook data access details are not fully specified in the extract.