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

This paper introduces a machine learning framework that utilizes a directed bipartite graph to model time-ordered predictive linkages between U.S. and Chinese equity markets, revealing a significant directional asymmetry where U.S. returns strongly predict Chinese intraday performance while the reverse effect is limited.

The Gist

Imagine you are trying to predict the weather in London (the Chinese stock market) by looking at the weather in New York (the U.S. stock market).

Here's the catch: When it's morning in London, it's the middle of the night in New York. When it's evening in New York, London is already asleep. They never see each other's weather at the same time.

This paper is about building a super-smart, automated weather forecaster that uses the "nightly weather report" from New York to guess what the "daytime weather" will be in London, and vice versa.

Here is the breakdown of how they did it, using simple analogies:

1. The Problem: Too Much Noise, Not Enough Signal

Financial markets are like a giant, chaotic party where thousands of people are shouting. Trying to predict which way a stock price will move is like trying to guess who will dance next in that crowd. It's noisy, confusing, and changes constantly.

Most people try to predict the future by looking only at the party they are currently in (the local market). This paper asks: "What if we listen to the party next door?"

2. The Solution: The "Night Watchman" Graph

The authors built a special tool called a Directed Bipartite Graph. Let's translate that into plain English:

• Bipartite: Imagine two separate groups of people. Group A is the New York stocks, and Group B is the Chinese stocks. They don't mix; they stand on opposite sides of a river.
• Directed: This means the information only flows one way. We are asking, "Does what Group A did yesterday tell us what Group B will do today?"
• The Graph: Think of this as a massive map of connections. The researchers drew lines (edges) between specific New York stocks and specific Chinese stocks only if there was a proven, statistical link between them.

How they drew the lines:
They didn't just guess. They used a "rolling window" (like looking through a moving window on a train). They checked: "Did Stock A in New York move in a specific way 24 hours ago, and did Stock B in China move in a matching way today?" If the answer was a strong "Yes," they drew a line connecting them. If the answer was "No" or "Maybe," they didn't draw a line.

This created a sparse, clean map that filters out the noise and only keeps the meaningful connections.

3. The Machine Learning "Brain"

Once they had this map of connections, they fed it into 10 different types of "brains" (Machine Learning models).

• Some brains were simple (like a basic calculator).
• Some were complex (like deep neural networks that learn patterns).
• Some were "Ensembles" (a committee of brains voting on the answer).

The goal was to see which brain could best use the "Night Watchman's" map to predict the next day's stock prices.

4. The Big Discovery: The One-Way Street

This is the most exciting part of the paper. They found a massive asymmetry (a one-way street):

• New York $\rightarrow$ China: The U.S. market is a crystal ball for China. Because the U.S. market closes before China opens, the U.S. has a full day of news, earnings reports, and economic data that China hasn't seen yet. The "Night Watchman" in New York knows a lot about what's coming. When the U.S. market moves, it sends a strong signal that helps predict the Chinese market the next day.
• China $\rightarrow$ New York: The reverse is weak. When China closes, the U.S. market is just waking up. By the time the U.S. opens, a lot of other global news has happened. The signal from China gets lost in the noise.

The Analogy:
Imagine New York is a lighthouse and China is a ship. The lighthouse (U.S.) shines a bright beam that the ship (China) can see clearly before it sets sail. But the ship's foghorn (China) is too quiet to be heard by the lighthouse keeper before he starts his day.

5. The Results: Why It Matters

The researchers tested their system and found:

• Better Predictions: Using the "Night Watchman" map (the graph) made the predictions much more accurate than just looking at the local market alone.
• The Best Combo: The most successful strategy was using U.S. closing prices to predict Chinese opening-to-closing prices.
• Profitability: When they simulated trading based on these predictions, the "Sharpe Ratio" (a score for how much profit you make per unit of risk) was high. This means the strategy was profitable and stable.

6. The Takeaway

This paper proves that in the global economy, information travels in waves. Because the U.S. and China trade at different times, the U.S. market acts as a "preview" for the Chinese market.

By using a structured map (the graph) to find exactly which stocks are connected, and then using AI to read that map, investors can see the future more clearly than by looking at the past alone. It's like having a secret shortcut that tells you what the next day's party will look like, based on how the party ended the night before.

Technical Summary

Here is a detailed technical summary of the paper "A Bipartite Graph Approach to U.S.-China Cross-Market Return Forecasting."

1. Problem Statement

The paper addresses the challenge of stock-level cross-market return prediction between the U.S. and Chinese equity markets. While machine learning (ML) has been widely applied to single-market forecasting, there is a lack of research on explicit stock-level prediction under realistic, non-overlapping trading session constraints.

Most existing cross-market studies focus on:

• Aggregate indices rather than individual stocks.
• Contemporaneous correlations or volatility spillovers rather than time-ordered predictive links.
• Symmetric relationships, ignoring the temporal causality imposed by different time zones.

The core problem is to construct a framework that leverages the non-overlapping trading hours (U.S. closes before China opens, and vice versa) to identify directed, predictive dependencies between specific U.S. and Chinese stocks, while maintaining economic interpretability and avoiding look-ahead bias.

2. Methodology

The authors propose a two-stage framework combining directed bipartite graph construction with ensemble machine learning.

A. Data and Definitions

• Data: Daily data from 2014–2021 for the top 500 stocks by market cap in the U.S. (NYSE/Nasdaq) and China (SSE/SZSE).
• Returns: The study uses market excess returns (raw return minus ETF return).
  • $pvCLCL$ (Previous Close-to-Close): Captures overnight information.
  • $OPCL$ (Open-to-Close): Captures intraday trading performance.
• Target: Predicting the $OPCL$ return of a target stock on day $t$.

B. Stage 1: Directed Bipartite Graph Construction

The authors construct a time-ordered bipartite graph $G = (V, E)$ where:

• Nodes: Two disjoint sets representing stocks in the Source Market ($X$) and Target Market ($Y$).
• Edges: Directed edges represent predictive relationships. An edge exists from $X_j$ to $Y_i$ if $X_j$'s past returns significantly predict $Y_i$'s future returns.
• Screening Mechanism:
  • A rolling window (250 days) is used.
  • For each pair $(X_j, Y_i)$, a univariate linear regression is performed: $y = \beta x + \alpha$.
  • Feature Selection: An edge is created if the absolute $t$-statistic of the slope coefficient $\beta$ exceeds a threshold $\tau = 2$ (approx. 5% significance).
  • Temporal Lag ($l$):
    • US $\to$ CN: $l=1$ (U.S. $t-1$ closes before China $t$ opens).
    • CN $\to$ US: $l=0$ (China $t$ closes before U.S. $t$ opens).
• Sparsification: The thresholding acts as a denoising mechanism, creating a sparse, interpretable feature set for the next stage.

C. Stage 2: Machine Learning Prediction

Using the selected edges as features, the framework forecasts returns using 10 different ML models:

1. Linear/Regularized: OLS, LASSO, Ridge Regression.
2. Kernel/Non-linear: Support Vector Machine (SVM) with RBF kernel.
3. Tree-Based/Boosting: XGBoost, LightGBM, Random Forests (RF), AdaBoost.
4. Ensembles: Average of predictions (ensemble-avg) and Median of predictions (ensemble-med).

The models are trained on a rolling window and evaluated strictly out-of-sample to prevent look-ahead bias.

3. Key Contributions

1. Novel Framework: Introduces a directed bipartite graph specifically designed for cross-market stock-level prediction, treating the graph as a sparse feature-selection layer rather than just a correlation network.
2. Exploitation of Time-Zone Asymmetry: Explicitly models the non-overlapping trading sessions to establish a clean causal ordering of information flow, distinguishing between $pvCLCL$ (overnight) and $OPCL$ (intraday) information.
3. Interpretability: Unlike "black-box" deep learning approaches, the graph structure provides an economically interpretable map of which specific stocks drive predictability in other markets.
4. Empirical Asymmetry Discovery: Systematically documents the directional asymmetry in information flow between the U.S. and Chinese markets at the individual stock level.

4. Key Results

A. Directional Asymmetry

• US $\to$ CN: U.S. previous-close-to-close ($pvCLCL$) returns contain substantial predictive information for Chinese intraday ($OPCL$) returns.
  • Performance: Sharpe Ratios (SR) frequently exceed 1.0, with some models approaching 2.0.
  • Best Models: Ridge Regression, LightGBM, and Ensemble methods performed best.
  • Feature Superiority: Using U.S. $pvCLCL$ (overnight) data outperformed using U.S. $OPCL$ (intraday) data, highlighting the value of overnight information transmission.
• CN $\to$ US: The reverse effect is significantly weaker.
  • Chinese returns have limited predictive power for U.S. stocks, resulting in much lower Sharpe Ratios across all models and quantiles.

B. Graph vs. Baselines

• Cross-Market vs. Same-Market: Models using cross-market graph features significantly outperformed:
  1. Non-graph baselines (using only same-market lagged returns).
  2. Graph-based same-market models (using graph structures within a single market).
• Ensemble Robustness: Ensemble methods (average and median) provided the most stable and robust performance, often matching or exceeding the best individual models.

C. Sensitivity Analysis

• Graph Structure: Randomly replacing edges in the graph led to a decline in performance, confirming that the predictive power relies on the economically meaningful structure of the graph, not just generic diversification.
• Temporal Decay: Increasing the time lag ($l$) between predictor and target reduced performance, confirming that cross-market predictive content decays rapidly over time.

5. Significance and Implications

• Economic Insight: The results reinforce the leading role of the U.S. market in global price discovery. The U.S. market acts as a primary information source for Chinese intraday trading, likely due to the U.S. market's size, liquidity, and the inclusion of global macroeconomic news in its overnight session.
• Methodological Advance: The paper demonstrates that structured machine learning (combining statistical screening with ML) can uncover complex, asymmetric dependencies that traditional econometric models or unstructured deep learning might miss.
• Practical Application: While the study focuses on predictive asymmetry rather than a deployable trading strategy (due to lack of transaction cost modeling), it provides a robust framework for:
  • Identifying specific cross-market hedging pairs.
  • Enhancing portfolio construction with foreign market signals.
  • Understanding information transmission channels between major global economies.

In conclusion, the paper successfully bridges the gap between network science and financial forecasting, proving that a time-ordered, bipartite graph approach effectively captures the asymmetric, stock-level predictive power of the U.S. market over the Chinese market.