Anomaly detection in European cryptocurrency exchange-traded products

Battle Count: 67

Highly relevant for quantitative trading in cryptocurrency ETP markets. The paper demonstrates that intraday anomalies are predictable one bar ahead (AUC up to 0.82), which could inform execution algorithms, risk management systems, and market-making strategies. The identification of microstructure correlates (effective spreads, Kyle's lambda, order flow imbalance) at anomaly bars is directly actionable. However, the very low precision (rarely above 0.10) limits direct trading signal utility. The cross-venue divergence anomaly is particularly relevant for statistical arbitrage and pairs trading strategies. The finding that short-term volatility and drawdown measures are more predictive than microstructure variables has practical implications for feature engineering in trading systems.

Moderate complexity. The anomaly definitions are clearly specified with explicit formulas and parameters (thresholds, windows, recovery criteria). The feature engineering involves 34 well-defined features across price/volatility and microstructure categories. The four classifiers are standard ML models with established implementations. The main complexity lies in: (1) proper implementation of the POT/GPD fitting with maximum likelihood estimation, (2) the rolling OLS regression for cross-venue divergence, (3) handling the asynchronous trading and active-bar filtering, (4) the chronological out-of-sample evaluation framework, and (5) managing severe class imbalance with class weights. No code or repository is provided.

The paper provides detailed methodology descriptions including formulas for all four anomaly definitions, feature construction (34 features), and model specifications. However, the data is obtained from a commercial provider (xyt) and is not publicly available. No GitHub repository is mentioned. Hyperparameters are stated to be fixed across instruments but exact values are not fully specified in the extract. The chronological 80/20 split methodology is clearly described.