Lead-Lag Relationships in Financial Markets: A Comparison of Multiple Clustering Algorithms

Battle Count: 80

Highly relevant to quantitative trading. The paper directly addresses lead-lag relationship detection for trading strategy construction, provides comprehensive backtesting with Sharpe ratios, maximum drawdown, and statistical significance tests. The comparison of clustering algorithms for identifying Leaders and Laggers is directly applicable to pairs trading and momentum strategies. The finding that MiniRocket-KMeans achieves Sharpe ratio of 0.866 with controlled drawdown is practically significant. However, the very large maximum drawdowns and marginal hit rates suggest practical implementation challenges.

Moderate complexity. The individual algorithms (KShape, MiniRocket, KMedoids) have established implementations in Python libraries (sktime, tslearn). However, the full pipeline requires: (1) sliding window time series extraction, (2) multiple clustering algorithm implementations, (3) DTW-based lead-lag matrix construction with mode/median estimation, (4) RowSum ranking for Leader/Lagger partition, (5) EWMA-based signal generation, and (6) backtesting framework. The ensemble voting mechanism and optimal K selection via silhouette coefficient add additional complexity. Computational efficiency is a key advantage of MiniRocket-KMeans over DTW-KMedoids.

The paper provides detailed algorithm descriptions, mathematical formulations, and experimental parameters (window size l=21, alpha=0.25, EWMA spans p={1,3,5,7}). However, no code repository is mentioned. The synthetic data generation process is well-described with the lagged multi-factor model. Real data uses Wharton's CRSP which is accessible but requires subscription. The 100 repetitions for synthetic experiments and specific parameter settings aid reproducibility.