We propose a directed bipartite-graph framework that screens time-ordered cross-market predictors (via rolling-window t-statistic thresholding) and feeds the sparse graph-selected features into a suite of machine-learning models to forecast next-session open-to-close returns; the approach uncovers a pronounced directional asymmetry — U.S. previous-close-to-close information is substantially more predictive for Chinese intraday returns than the reverse — and combining graph-based selection with cross-market features materially improves out-of-sample Sharpe Ratios across many models .
速读内容
Directed bipartite-graph construction and screening
Two disjoint node sets represent source (e.g., U.S.) and target (e.g., China) markets; edges are selected when the rolling-window regression t-statistic |t_beta| > τ (τ = 2 in experiments), producing a sparse, time-ordered biadjacency matrix used for feature selection rather than formal inference .
Key parameters: look-back window w = 250 trading days, update/retrain every 10 days, selection threshold τ = 2, lag l = 1 for US→CN and l = 0 for CN→US to preserve non-overlapping trading-session ordering .
Cross-market connectivity: stock- and sector-level patterns
Biadjacency heatmaps show non-block-diagonal structure: predictive links span sectors, indicating economically interpretable cross-sector transmission channels rather than only within-sector links .
Aggregated sector-level median absolute t-statistic reveals particular cross-sector strengths (e.g., Chinese financials showing links to U.S. utilities), highlighting systematic sectoral dependencies .
Prediction framework and models
For each target stock Yi, predictors are the n source stocks selected via the graph for day t−l; forecasting function: r_Yi^(t) = F_i(r_X1^(t-l),...,r_Xn^(t-l); θ) trained in rolling windows to ensure strict out-of-sample evaluation .
Models used: OLS, LASSO, RIDGE, SVM (RBF), XGBoost, LGBM, Random Forest, AdaBoost, ensemble-avg and ensemble-med; hyperparameter tuning done within rolling training windows .
Main economic findings — directional asymmetry and model performance
U.S. pvCLCL → Chinese OPCL forecasting yields materially higher Sharpe Ratios than the reverse direction (CN→US), demonstrating pronounced informational asymmetry where U.S. previous-close information is more informative for Chinese intraday returns .
Performance overview: most models produce SR > 1 when forecasting CN using US predictors; RIDGE, LGBM and ensemble methods are consistently strong (some SRs approach or exceed 2.0 in top signal quantiles), while SVM underperforms in many settings .
PnL/backtest behavior: cumulative PnL trajectories for top quantiles show persistent upward slopes in the US→CN setting, supporting economic significance of predictive signals (pre-cost, stylized sizing) .
Quantitative strategy mechanics (graph + ML)
Feature selection: for each Yi pick top-n source stocks by |t_beta| from rolling regressions; preserve temporal ordering (lag l) to avoid look-ahead bias .
Portfolio formation: daily ranking by |predicted return| and nested quantile portfolios (qr1..qr6); position sizing uses capped liquidity proxy b_i^(t) = min(0.001×MDV_i^(21), L) with L set per-market (USD/CNY) .
Typical hyperparams: training window w=250, retrain every 10 days, lookback features for same-market baseline use past 25 days; experiments compare graph-based cross-market, graph-based same-market, and non-graph baselines .
Robustness and sensitivity
Feature-replacement test: randomly replacing an increasing fraction of graph edges (while preserving in-degree) degrades SRs — predictive gains depend on economically meaningful graph structure rather than random links .
Temporal decay: increasing lag l (using older predictors) reduces SRs — recency matters; highest-magnitude signals (top quantile) are relatively more robust to lagging .
Practical caveats
Results are pre-transaction-cost and use stylized position sizing (no market-impact or capacity optimization); graph screening involves multiple testing risks and may include spurious edges — interpret SRs as evidence of predictive asymmetry, not deployable alpha without further trading-level engineering .
总体两阶段流程:第一步在滚动训练窗口内对所有有序跨市场股票对进行成对单变量线性回归并基于 t 统计量筛选显著连接从而构建“有向二部图”;第二步以每个目标股票的入度所对应的若干源市场股票的当期(lagged)回报作为特征输入十种机器学习模型进行预测并滚动检验。作者把图构建定位为“稀疏化、可解释”的特征选择层而非严格的因果推断步骤。, ,
回归筛选细节:对每对 (X_j, Y_i) 在滚动窗口 w 内用简单线性回归 y ~ x,计算斜率的 t 统计量 t_beta 并比较阈值 τ(实验中 τ=2)来决定是否保留边;滞后参数 l 用以保证源市场在目标市场交易前完全可观测(例如预测 CN 用 US 信息时 l=1,反方向 l=0)。作者明确指出这是“边对边的边际筛选”,存在多重检验与伪边的风险,但其目标是出于预测性能而非对单条边做正式统计检验。, , ,