Explainable Patterns in Cryptocurrency Microstructure

Battle Count: 68

Highly relevant to quantitative trading. The paper directly addresses short-horizon return prediction using order book microstructure features, validates tradability through realistic taker and maker backtests, demonstrates cross-asset portability of feature libraries (reducing engineering costs), and provides actionable insights on adverse selection risk during extreme events. The SHAP-based interpretability connects model predictions to economic theory, enabling risk controls (e.g., deactivating signals in wide-spread regimes). The flash crash analysis provides critical stress-testing evidence for strategy robustness.

Moderate complexity. The core modeling pipeline (CatBoost + SHAP + Optuna) uses well-established libraries. However, the full implementation requires: (1) processing sub-second order book and trade data from Binance API, (2) constructing synchronized time series with forward-filling, (3) engineering relative features (spreads-to-mid, VWAP-to-mid deviations, order flow imbalance), (4) implementing walk-forward cross-validation with purging, (5) custom GMADL objective function, (6) conservative taker/maker backtest with pessimistic inventory marking, and (7) cross-asset SHAP comparison analysis. The data volume (1-second frequency over ~3.5 years for 5 assets) and real-time simulation aspects add engineering challenges.

The paper describes a clear methodology with specific data source (Binance Futures), feature engineering details, model configuration (CatBoost with Optuna hyperparameter optimization), and evaluation framework. However, no code repository is mentioned, and the exact feature engineering code, hyperparameter ranges, and backtest implementation details are not fully specified. The data is from a public exchange but requires significant preprocessing. The SHAP analysis and backtest logic are described but not provided as code.