DeltaLag introduces the first end-to-end deep learning framework for discovering and exploiting dynamic lead-lag relationships in financial markets. Unlike traditional two-step statistical approaches that assume persistent lead-lag patterns, DeltaLag uses a sparsified cross-attention mechanism to adaptively identify, for each target stock, its top-k leading stocks and their corresponding optimal lag values on a daily basis. The model encodes historical price-volume features via a temporal encoder (e.g., LSTM), computes attention scores between the target's query and candidate leaders' key sequences over a range of lags, and selects the top-k pairs. The raw features of the identified leaders at their lagged times are then aggregated via attention weights and fed into an MLP to predict the target's next-day return. A monotonic logistic ranking loss is employed to optimize cross-sectional stock rankings directly, enhancing trading performance in long-short portfolios.

Empirical evaluations on US equities from 2010 to 2023 across S&P 500, NASDAQ, and NYSE markets demonstrate that DeltaLag significantly outperforms fixed-lag and self-lead-lag baselines, confirming the importance of dynamic, pair-specific lags and cross-asset dependencies. It also surpasses methods relying on precomputed lead-lag graphs from statistical correlations, highlighting the weak momentum of lead-lag relationships and the need for adaptive detection. Furthermore, DeltaLag achieves superior cumulative returns and Sharpe ratios compared to a wide range of temporal and spatio-temporal deep learning models, while offering enhanced interpretability by explicitly modeling cross-stock lead-lag structures. The framework's ability to extract economically meaningful signals directly from raw price-volume data underscores its practical value for quantitative investment.