A Bipartite Graph Approach to U.S.-China Cross-Market Return Forecasting

Battle Count: 150

Highly relevant to quantitative trading. The paper demonstrates economically meaningful cross-market return predictability with Sharpe Ratios exceeding 1 (up to ~1.8) for forecasting Chinese stocks using U.S. information. The directional asymmetry (U.S.→China >> China→U.S.) aligns with established literature on U.S. market leadership. The graph-based feature selection provides interpretable transmission channels. However, results are pre-cost and stylized, limiting direct implementation. The methodology is applicable to overnight information trading strategies exploiting non-overlapping sessions. Sector-level patterns (technology, consumer defensive showing stronger effects) provide actionable insights.

Moderate to high complexity. The two-stage pipeline (graph construction via rolling-window t-tests + ML prediction) requires careful temporal alignment across non-overlapping trading sessions. Graph construction involves pairwise regressions across 500×500 stock pairs per rolling window. Multiple ML models with hyperparameter tuning are needed. The rolling-window scheme (250-day window, 10-day update) adds computational overhead. However, the methodology is well-described and uses standard statistical and ML tools. The main complexity lies in correct temporal alignment, handling different trading calendars, and managing the large-scale pairwise screening.

The paper provides detailed methodology, hyperparameter settings (Table A.1), data sources (CRSP for U.S., Wind for China), and experimental setup (250-day window, 10-day update, 500 stocks per market). However, no code repository is mentioned, and the Wind database is a commercial data source. The universe selection uses full-sample market capitalization (2014-2021), introducing a look-ahead component acknowledged by the authors. Reproducibility is moderate given the detailed descriptions but limited by data access and absence of code.