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

Jing Liu, Maria Grith, Xiaowen Dong, Mihai CucuringuarXiv 2026 · arXiv

TL;DR

This paper introduces a structured machine learning framework that leverages the non-overlapping trading hours of the U.S. and Chinese equity markets to forecast stock-level returns. By constructing a time-ordered, directed bipartite graph via rolling-window hypothesis testing, the authors create an interpretable feature-selection layer for downstream machine learning models. Their empirical findings reveal a striking directional asymmetry: U.S. market returns contain substantial predictive power for subsequent Chinese intraday returns, whereas the reverse effect is highly limited.

What Problem Does This Paper Address?

Predicting stock returns is notoriously difficult due to low signal-to-noise ratios, market non-stationarity, and complex nonlinear dependencies. While machine learning has become a staple in single-market forecasting, cross-market stock-level prediction remains largely unexplored under realistic trading-session constraints.

Most prior cross-market literature operates at the aggregate index level or focuses on contemporaneous co-movements and spillovers rather than explicit, stock-to-stock return forecasting. The authors address this gap by exploiting a unique structural feature of global finance: the U.S. and Chinese equity markets have completely non-overlapping trading hours. This asynchronous schedule provides a clean, time-ordered window to test whether information from one market can actively predict individual stock returns in the other.

Methods

The authors propose a novel two-stage framework that blends econometric screening with machine learning:

  1. Directed Bipartite Graph Construction (Feature Selection): The markets are treated as two disjoint sets of vertices (source market $\mathcal{X}$ and target market $\mathcal{Y}$). For every ordered pair of U.S. and Chinese stocks, the authors run a univariate linear regression of target returns on lagged source returns over a 250-day rolling training window. Edges are formed only when the regression coefficient's $t$-statistic exceeds an absolute threshold ($\tau = 2$). This acts as a computationally tractable, data-driven sparsification layer to filter out noise and isolate meaningful cross-market predictive links.

  2. Downstream Machine Learning Forecasting: The selected top-$n$ predictive stocks from the graph serve as features for ten different downstream forecasting models. The model suite spans linear methods, regularized models, tree-based architectures, and ensembles:

  3. Ordinary Least Squares (OLS)

  4. LASSO & Ridge Regression
  5. Support Vector Machines (SVM)
  6. XGBoost & LightGBM (LGBM)
  7. Random Forests (RF) & AdaBoost
  8. Ensemble averages and medians

The target variable is the next-session open-to-close (OPCL) return, and models are updated every 10 days to prevent look-ahead bias.

[Source Market (Day t-1)] ---> (Statistical Screening Layer) ---> [Target Market (Day t)]
      (U.S. Stocks)               (|t-statistic| > 2)               (Chinese Stocks)

Datasets

The study utilizes daily stock data spanning from 2014 through 2021:

  • U.S. Market: Sourced from the Center for Research in Security Prices (CRSP), focusing on the New York Stock Exchange (NYSE) and Nasdaq.
  • Chinese Market: Sourced from the Wind Database, covering the Shanghai Stock Exchange (SSE) and Shenzhen Stock Exchange (SZSE).

Universe Selection & Benchmarks: The authors select the 500 largest stocks by average market capitalization from each country. To isolate stock-specific signals, they evaluate market excess returns by subtracting the returns of representative Exchange-Traded Funds (ETFs): SPY for the U.S. and 513500.SH for China.

Key Insights and Ideas

  • Pronounced Directional Asymmetry: The flow of predictive information is heavily one-sided. While U.S. previous-close-to-close (pvCLCL) returns strongly predict Chinese intraday returns, Chinese returns offer negligible predictive power for U.S. stocks.
  • The Power of Overnight Transmission: U.S. previous-close-to-close (pvCLCL) returns serve as significantly better predictors for Chinese intraday returns than U.S. intraday (OPCL) returns, highlighting that information compounds and transmits effectively during the overnight gap.
  • Beyond Block-Diagonal Structures: Aggregating the bipartite graph at the sector level reveals that cross-market linkages are not merely confined to within-sector pairs (e.g., U.S. Tech predicting Chinese Tech). Instead, complex cross-sector channels emerge—such as the U.S. utilities sector showing strong predictive links to the Chinese financial services sector.

Key Results

  • High Risk-Adjusted Returns: When predicting Chinese stocks using U.S. predictors, regularized models (Ridge), tree-based models (LightGBM), and ensemble methods achieved strong out-of-sample performance, with annualized Sharpe Ratios (SRs) exceeding 1.0 across most portfolio quantiles, and some approaching 2.0 in top-ranked portfolios.
  • Cross-Market vs. Same-Market: The graph-based cross-market approach consistently outperformed both the non-graph baselines and the graph-based same-market baselines, proving that foreign market lag contains distinct, incremental alpha.
  • Sensitivity and Decay: In feature-randomization tests, performance steadily degraded as the true graph structure was perturbed, confirming that the model relies on economically meaningful linkages rather than generic diversification. Furthermore, increasing the temporal lag ($l \ge 2$ days) caused a sharp drop in Sharpe Ratios, proving that cross-market predictive alpha decays rapidly over time.

Limitations and Future Work

The authors openly acknowledge several caveats in their experimental setup:

  • Look-Ahead Component in Universe Selection: The 500-stock universe was selected based on full-sample average market capitalization (2014–2021). While practical for ensuring liquidity, a truly investable backtest would require time-$t$ local universe reconstitution.
  • Stylized Trading Design: The evaluation abstracts away from real-world frictions such as transaction costs, short-sale constraints, market impact, and liquidity-optimized portfolio weighting. The results demonstrate predictive strength rather than a production-ready execution strategy.
  • Spurious Edges: Large-scale pairwise screening without multiple-testing corrections (like False Discovery Rate adjustments) likely introduces some spurious edges in finite samples.

Future Directions: The authors suggest expanding the framework to European and other Asian markets, applying Graph Neural Networks (GNNs) directly to the bipartite graph to capture nonlinear structural dependencies, and exploring time-series-specialized Large Language Models (LLMs) for cross-market analysis.

Summary updated

Abstract

This paper studies cross-market return predictability through a machine learning framework that preserves economic structure. Exploiting the non-overlapping trading hours of the U.S. and Chinese equity markets, we construct a directed bipartite graph that captures time-ordered predictive linkages between stocks across markets. Edges are selected via rolling-window hypothesis testing, and the resulting graph serves as a sparse, economically interpretable feature-selection layer for downstream machine learning models. We apply a range of regularized and ensemble methods to forecast open-to-close returns using lagged foreign-market information. Our results reveal a pronounced directional asymmetry: U.S. previous-close-to-close returns contain substantial predictive information for Chinese intraday returns, whereas the reverse effect is limited. This informational asymmetry translates into economically meaningful performance differences and highlights how structured machine learning frameworks can uncover cross-market dependencies while maintaining interpretability.
BibTeX
@misc{liu2026bipartite,
  author = {Jing Liu and Maria Grith and Xiaowen Dong and Mihai Cucuringu},
  title = {{A Bipartite Graph Approach to U.S.-China Cross-Market Return Forecasting}},
  year = {2026},
  eprint = {2603.10559},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.10559},
}
Figure