Reproduce “DeltaLag: Learning Dynamic Lead-Lag Patterns in Financial Markets” on OpenResearch

About this paper

Wanyun Zhou, Saizhuo Wang, Mihai Cucuringu, Zihao Zhang, Xiang Li, Jian Guo, Chao Zhang, Xiaowen Chu

DeltaLag, an end-to-end deep learning method developed by researchers across several institutions, identifies and exploits dynamic lead-lag patterns in financial markets by dynamically learning pair-specific lag values. This framework achieved superior cumulative returns, approximately 10 basis points daily, compared to state-of-the-art baselines on US equities data.

Key insights

• Lead-lag relationships in financial markets are highly dynamic and non-stationary, requiring adaptive models that can continuously learn and adjust to transient patterns rather than relying on historically stable correlations.
• Dynamically learned pair-specific lag values and cross-asset dependencies provide more informative predictive signals for stock returns than fixed-lag assumptions or a stock's own historical patterns.
• An end-to-end deep learning framework, optimized with a ranking-based loss function, yields significantly better performance for portfolio construction by jointly learning lead-lag detection and signal generation.

Summary by alphaXiv.