A Bipartite Graph Approach to U.S.-China Cross-Market Return Forecasting
Authors: Jing Liu, Maria Grith, Xiaowen Dong, Mihai Cucuringu
Organizations: Department of Statistics, University of Oxford, UK · Finance Department, Neoma Business School, France · Department of Engineering Science, University of Oxford, UK · Department of Mathematics, University of California Los Angeles, US · Oxford-Man Institute of Quantitative Finance, University of Oxford, UK
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.
Figures & tables
Figure 1: Schematic illustration of the directed bipartite graph linking source-market stocks to target-market stocks based on significant predictive relationships.
Figure 2: Example time series of U.S. pvCLCL returns for CDNS and Chinese OPCL returns for 002410.XSHE over the rolling training window. The series are shown for illustrative purposes to highlight cross-market co-movement underlying the detected predictive link.
Figure 3: Heatmap of the directed biadjacency matrix for a representative trading day. Rows correspond to Chinese stocks and columns to U.S. stocks, grouped by sector. Each entry represents the t-statistic from the rolling-window regression of Chinese returns on lagged U.S. returns. Colour intensity reflects the magnitude and sign of the predictive relationship.
Figure 4: Sector-level heatmap of the absolute median t-statistic in the time-averaged biadjacency matrix of the directed cross-market graph. Rows correspond to Chinese sectors and columns to U.S. sectors. Each entry reports the median absolute predictive strength across all stock pairs within the corresponding sector-by-sector block.
Figure 5: The figure shows the 25th, 50th, and 75th percentiles of the in-degree distribution of target nodes by day. US-CN represents the number of U.S. pvCLCL nodes selected to predict Chinese OPCL returns, while CN-US represents the number of Chinese pvCLCL nodes selected to predict U.S. OPCL returns.
Figure 6: Schematic illustration of the rolling training and prediction framework. For each target stock Yi , returns over the look-back window [t−w,t−1] are regressed on lagged source-market returns over [t−l−w,t−l−1] . The bottom row represents the out-of-sample prediction of rYi(t) using source returns observed at t−l , thereby preserving temporal ordering and eliminating look-ahead bias.
Figure 7: Timeline of opening and closing times for the U.S. and Chinese stock markets. The non-overlapping trading sessions induce a natural temporal ordering of information, with U.S. day t−1 close preceding Chinese day t trading, and Chinese day t close preceding U.S. day t trading.
Figure 8: Sharpe Ratios for forecasting Chinese OPCL returns using U.S. pvCLCL and OPCL returns as predictors.
Figure 9: Sharpe Ratios for forecasting U.S. OPCL returns using Chinese returns as predictors.
Figure 10: Cumulative daily Profit and Loss (PnL) from forecasting Chinese OPCL returns using U.S. pvCLCL returns as predictors. Each panel corresponds to a different machine learning model, and coloured curves represent nested quantile portfolios ranked by the absolute value of predicted returns.
Figure 11: Sharpe Ratios for forecasting Chinese OPCL returns using Chinese returns as predictors.
Figure 12: Comparison of Sharpe Ratios across graph-based cross-market, graph-based same-market, and non-graph-based baseline specifications, using pvCLCL returns as predictors. CN-CN denotes using Chinese stocks to forecast Chinese stocks, while US-CN denotes using U.S. stocks to forecast Chinese stocks.
Figure 13: Performance differentials (Sharpe Ratio deltas) relative to the non-graph-based baseline, using pvCLCL returns as predictors. US-CN denotes using U.S. stocks to forecast Chinese stocks, while CN-CN denotes using Chinese stocks to forecast Chinese stocks.
Figure 14: Median forecasting performance under graph randomization (a) and increasing temporal lag (b). Panel (a) reports Sharpe Ratios as a function of the fraction of replaced edges while preserving in-degree. Panel (b) reports Sharpe Ratios as the lag parameter l increases, measuring the effect of input recency.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B.1 : Comparison of Sharpe Ratios across graph-based cross-market, graph-based same-market, and non-graph-based baseline specifications, using OPCL returns as predictors. US-CN denotes using U.S. stocks to forecast Chinese stocks, while CN-CN denotes using Chinese stocks to forecast Chinese stocks.
Figure B.2 : Performance differentials (Sharpe Ratio deltas) relative to the non-graph-based baseline, using OPCL returns as predictors. US-CN denotes using U.S. stocks to forecast Chinese stocks, while CN-CN denotes using Chinese stocks to forecast Chinese stocks.
Figure C.1 : Sharpe Ratios of forecasting models conditional on SPY absolute pvCLCL return nested quantiles, measuring predictive performance following different magnitudes of U.S. market shocks.
Figure D.1 : Sector-level Sharpe Ratios for cross-market forecasting (using U.S. pvCLCL returns to forecast Chinese OPCL returns).
Figure E.1 : Sharpe Ratios for forecasting Chinese OPCL returns when the predictor set includes both U.S. and Chinese stocks. Panels vary the combination of pvCLCL and OPCL features across the two markets.
Financial forecasting uses market, company, and economic data to predict prices, returns, volatility, and business outcomes for decisions and risk management.
Accurate prediction of equity returns remains a major challenge in computational finance due to the non-stationary, nonlinear, and low signal-to-noise ratio nature of financial time series. This paper proposes a hybrid two-stage architecture that combines a long short-term memory (LSTM) network with an XGBoost gradient-boosted regressor for multi-horizon stock return prediction across a diversified panel of 14 U.S. equities spanning six industry sectors. The LSTM component, comprising two stacked layers with 64 hidden units, processes 60-day sliding windows of five sequential market features to produce 64-dimensional temporal embeddings that encode learned sequential market dynamics. These embeddings are concatenated with 14 hand-crafted technical indicators to form a 78-dimensional hybrid feature vector, which is subsequently passed to an XGBoost regressor tuned via 3-fold cross-validation grid search. The framework is trained on a multi-stock pooled corpus using strict chronological splits and per-stock MinMaxScaling to prevent look-ahead bias, and evaluated across four prediction horizons of 30, 90, 252, and 365 trading days. Experimental results demonstrate that the hybrid model achieves a test RMSE of 0.0949 on the 30-day horizon, roughly one-third that of the standalone LSTM baseline, while marginally matching or surpassing the XGBoost-Only baseline across the majority of stocks. Directional accuracy rises with horizon length, reaching 97.6% at 365 days; we show, however, that this largely tracks the high base rate of positive long-horizon returns in the sample, and we therefore benchmark directional accuracy against a naive always-positive predictor and treat the above-base-rate gap at short horizons as the more informative signal. A composite investment scoring framework derived from multi-horizon predictions is further proposed to support portfolio ranking and decision support.
This paper introduces a graph-theoretic approach for predicting market regimes in foreign exchange (FX) currency prices. Specifically, the proposed model incorporates exogenous macroeconomic variables to update localized node features via message-passing operations. Utilizing the Graph Tsetlin Machine (GraphTM) framework, we empirically demonstrate the efficacy of this approach in anticipating market regimes for the US Dollar and Japanese Yen currency pair (USD/JPY). By representing multivariate macroeconomic drivers and technical indicators as hypervectorized directed multigraphs, the GraphTM leverages structured message passing to construct deep, interpretable logical clauses capable of recognizing complex sub-graph patterns.
Christian Blakely, Melanie Gilmore
†Centre for Artificial Intelligence Research, University of Agder, Norway · ‡Head of AI, Bernoly AG, Zurich, Switzerland · ¶La Jolla Private Wealth Group, Wells Fargo Advisors, La Jolla, CA, USA
We present an interpretable machine learning pipeline to decompose Cross-Sectional Equity Return Predictability into auditable factor contribution. We apply an XGBoost model with TreeSHAP attribution and conduct stress testing on 3632 Chinese A-share stocks from 2009 until 2019. Using 60-month, rolling windows over 55 months of out-of-sample data, XGBoost obtains a mean AUC of 0.547 and +2.38%/month (Newey-West t = 5.94; Annualized Sharpe 2.23) long-short spread for the top vs bottom quintiles. This alpha is persistent after adjusting for the Carhart four-factor model (+2.31%/month; t = 7.48). SHAP Decomposition indicates that behavioral signals (turnover and momentum) account for 58.2% of predictive attribution compared to 10.7% for valuation ratios, on average, across 55 industry groups. Ablation analysis serves to cross-validate this ranking and provides evidence that SHAP and ablation diverge in a manner that highlights feature substitutability structure that is largely invisible to either method used in isolation.
Xiao Han, Yao Xiao, Zhen Zhang +1
Goizueta Business School Emory University Atlanta, GA, USA · College of Computing Georgia Institute of Technology Atlanta, GA, USA · School of Data Science University of Pennsylvania Philadelphia, PA, USA +1