Intraday Market Return Predictability in A-Shares: A Factor Zoo & Machine Learning Approach

📌 Executive Summary

This repository presents a high-fidelity replication and localized extension of the landmark study "Intraday Market Return Predictability Culled from the Factor Zoo" (Aleti, Bollerslev, & Siggaard, 2025, Management Science).

Focusing on the China A-Share market, this project explores the time-series predictability of the HS300 ETF by leveraging a "zoo" of 105 high-frequency characteristic-sorted portfolios. By integrating financial econometrics (Jump-Diffusion decomposition) with a robust Machine Learning ensemble (Linear, Tree-based, and Deep Learning models), this research demonstrates significant out-of-sample alpha even after accounting for realistic transaction costs and market frictions.

---

🛠 Methodology & Innovation

1. High-Frequency Factor Zoo (China-Specific)

Unlike the original study's US focus, this project constructs 105 high-frequency anomalies specifically tailored for the A-share market (following Wang & Zhu, 2024). These factors span six dimensions:

• Momentum & Volatility: Capturing short-term trend persistence and risk premia.
• Trading Friction: dtv (Volume), tur (Turnover), and iv (Idiosyncratic Volatility).
• Fundamental Categories: Value-versus-growth, Investment, and Profitability.

2. Return Decomposition: Continuous vs. Jump

To enhance signal-to-noise ratios, market returns ($r_{t}$) are decomposed into continuous price dynamics and jump components using high-frequency econometric filters: $$r_{t} = r_{cts, t} + r_{jmp, t}$$ The inclusion of the jump component ($r_{jmp}$) as a predictive feature significantly improves model robustness during volatile market regimes.

3. Predictive Model Zoo (Extended)

To capture both linear dependencies and complex non-linear interactions, we implement a diverse suite of models:

• Regularized Linear Models: Ridge and Elastic Net (Enet) for high-dimensional feature selection and shrinkage.
• Ensemble Tree Models: Random Forest (RF) and Gradient Boosting Regression Trees (GBRT) to capture non-linear factor interactions and structural breaks.
• Deep Learning: A Multi-layer Feed-forward Neural Network (FNN) designed to learn high-level representations of intraday liquidity patterns.

---

📈 Backtest Performance (HS300 ETF)

The strategy employs a Signal-Quantile Grid approach, executing trades on the HS300 ETF at 5-minute intervals based on the top 5% signal thresholds.

Strategy Highlights:

• Execution Logic: Next-bar Open price execution with realistic slippage modeling.
• Cost Sensitivity: Incorporated 0.85 bps (Buy) and 5 bps (Sell/Short) transaction fees.
• Alpha Generation: The Ensemble-All+Cts model achieves a cumulative return of 7.695 with an average net profit per trade of 0.085%.

Metric	Performance (Ensemble Model)
Cumulative Return	7.695
Max Drawdown	24.8%
Total Trades	9,032
Avg. Net Profit/Trade	0.085% (After-cost)

---

🔍 Interpretability & SHAP Analysis

Using SHAP (SHapley Additive exPlanations), we deconstruct the model's decision-making process to ensure economic validity.

Key Insights:

• Liquidity Proxies: Trading volume (dtv) and Share price (pps) emerge as the most influential predictors, supporting the Slow-Moving Capital hypothesis.
• Risk Dynamics: Idiosyncratic Volatility (iv) provides critical predictive power during market opening and closing auctions.
• Microstructure: The predictability is heavily concentrated in the first 30 minutes of trading, reflecting information asymmetry resolution in A-shares.

---

✉️ Contact

[Yilin Yang] Quantitative Researcher Candidate

Email: yang13515360252@163.com