Paper: arXiv 2507.07107
Authors: Yimin Du
Abstract
This paper presents a comprehensive machine learning framework for quantitative trading that achieves superior risk-adjusted returns through systematic factor engineering, real-time computation optimization, and cross-sectional portfolio construction. Our approach integrates multi-factor alpha discovery with bias correction techniques, leveraging PyTorch-accelerated factor computation and advanced portfolio optimization. The system processes 500-1000 factors derived from open-source alpha101 extensions and proprietary market microstructure signals. Key innovations include tensor-based factor computation acceleration, geometric Brownian motion data augmentation, and cross-sectional neutralization strategies. Empirical validation on Chinese A-share markets (2010-2024) demonstrates annualized returns of $20%$ with Sharpe ratios exceeding 2.0, significantly outperforming traditional approaches. Our analysis reveals the critical importance of bias correction in factor construction and the substantial impact of cross-sectional portfolio optimization on strategy performance. Code and experimental implementations are available at: https://github.com/initial-d/ml-quant-trading
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents advanced mathematical frameworks including tensor operations, geometric Brownian motion augmentation, and complex portfolio optimization formulations, while demonstrating high empirical rigor with detailed performance metrics, backtest data over a 14-year period, and publicly available code implementations.
Research Flowchart
flowchart TD A["Research Goal: ML-Enhanced Quant Trading with<br>Factor Optimization & Bias Correction"] --> B["Data & Inputs<br>1000+ Factors: Alpha101 + Microstructure<br>Chinese A-Shares (2010-2024)"] B --> C["Methodology: Tensor-Accelerated<br>Factor Computation via PyTorch"] C --> D["Processing: Cross-Sectional<br>Portfolio Optimization & Neutralization"] D --> E["Innovation: Bias Correction<br>+ Geometric Brownian Motion Augmentation"] E --> F["Outcomes: 20% Annualized Return<br>Sharpe Ratio > 2.0"] F --> G["Key Finding: Bias Correction &<br>Portfolio Optimization Critical to Performance"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a mask‑first design for rolling‑window factor pipelines and a full open‑source quantitative trading system that includes a 213‑factor GPU engine, an Adjusted‑MSE loss, block‑bootstrap GBM augmentation, and Markowitz‑Ledoit‑Wolf portfolio optimisation.
Source sentence
We resolve this with a mask-first design: a Boolean tradability mask is constructed at data load time and threaded through every operator, so that no window ever reads a non-tradable price.
Assumes 6
- Daily price‑move limits (+/-10% main‑board, +/-20% STAR/ChiNext) make some closing prices non‑executable. market
Source sentence
Rolling-window factor pipelines for Chinese A-share markets contain a subtle but costly flaw: daily price-move limits (+/-10% main-board, +/-20% STAR/ChiNext) render a fraction of closing prices non-executable
- Only trades on days where both current and next‑day masks are 1 are used for training. data
Source sentence
A sample enters the dataset only if both Mt,i = 1 and Mt+1,i = 1—this bidirectional check is critical because even if factors are clean at t, if the stock hits a limit at t+1, the label return is non-executable.
- Synthetic returns are generated from per‑stock drift and volatility estimated on masked history and follow a log‑normal GBM with block‑bootstrap sampling. model
Source sentence
Generate synthetic log-returns: ˜rt,i = ˆµi + ˆσi · ϵt where innovations ϵt ∼N(0, 1) are drawn in 21-day blocks (one calendar month) to preserve short-range serial correlation.
- Transaction costs are modeled as a linear cost of 5–8 basis points per unit of turnover. execution
Source sentence
Transaction costs.. Net returns deduct a linear cost: rnet_t = w⊤_{t‑1}rt −c·∥wt −wt‑1∥1 with c = 5–8 bps per unit of turnover.
Show 2 more
- The strategy is constrained to long‑only positions due to A‑share short‑selling restrictions. market
Source sentence
Long-only constraint.. The system is constrained to long-only due to A-share short-selling restrictions.
- The linear transaction‑cost model is optimistic for large portfolios where market impact is convex. execution
Source sentence
Our 5–8 bps linear cost model is optimistic for large portfolios where market impact is convex in trade size.
Data
2022 – 2024 · Chinese A-share stocks · equities · daily
- Synthetic panel synthetic public
Source sentence
Synthetic panel.. 3 000 stocks × 3 500 days (≈14 years from 2010-01-04), single-factor GBM with market β = 0.55, annual drift 7 %, volatility 32 %, halt probability 1 %, ±10 % price limit.
- Real A-share panel market-prices commercial
Source sentence
Real A-share panel.. Daily OHLCV from Tushare, January 2015–December 2024, average ≈3 200 active stocks after masking.
Costs
Modeled
Linear cost of 5–8 bps per unit of turnover is subtracted from net returns.
Source sentence
Transaction costs.. Net returns deduct a linear cost: rnet_t = w⊤_{t‑1}rt −c·∥wt −wt‑1∥1 with c = 5–8 bps per unit of turnover.
Finds 5
Sharpe 1.63 · annual return 20.0%
- The mask‑first contract alone adds 0.44 Sharpe points, the largest single component contribution. performance
Source sentence
Ablation shows the mask contract is the single largest contributor (+0.44), exceeding any model or loss choice.
- Full system achieves annualised Sharpe 2.05 on synthetic data and 1.63 on real A‑share data (2022–2024). performance
Source sentence
The system achieves Sharpe 2.05 on synthetic data and 1.63 on real A-share data (2022–2024), with the mask contract alone accounting for 0.44 Sharpe points.
- Adjusted‑MSE loss contributes an additional 0.27 Sharpe points. performance
Source sentence
The Adjusted‑MSE loss adds 0.27, GBM augmentation adds 0.19, and Ledoit–Wolf shrinkage adds 0.18.
- Block‑bootstrap GBM augmentation adds 0.19 Sharpe points, especially benefiting the Transformer model. performance
Source sentence
The Transformer benefits more from GBM augmentation (+0.19 Sharpe) than the MLP (+0.12).
- Upstream contamination (using non‑executable prices) inflates IC by 18% and reduces realised Sharpe by 0.44 points. predictive-accuracy
Source sentence
On real A-share data it inflates apparent information coefficient by 18% while reducing realised Sharpe by 0.44 points, because the model learns to predict returns it cannot trade.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (76)
- Empirical asset pricing via machine learning (2020)
- Deep learning in asset pricing (2024)
- Dissecting characteristics nonparametrically (2020)
- Capital asset prices: A theory of market equilibrium (1964)
- The valuation of risk assets (1965)
- The arbitrage theory of capital asset pricing (1976)
- Common risk factors in the returns on stocks and bonds (1993)
- A five-factor asset pricing model (2015)
and 68 more
- On persistence in mutual fund performance (1997)
- Returns to buying winners and selling losers (1993)
- … and the cross-section of expected returns (2016)
- The characteristics that provide independent information about average US monthly stock returns (2017)
- 101 formulaic alphas (2016)
- Characteristics are covariances: A unified model of risk and return (2019)
- Taming the factor zoo: A test of new factors (2020)
- Does academic research destroy stock return predictability? (2016)
- Bond risk premiums with machine learning (2021)
- Does it pay to follow anomalies research? Machine learning approach with international evidence (2021)
- Machine learning in the Chinese stock market (2022)
- Deep learning for predicting asset returns (2023)
- Factor investing in Chinese A-share markets (2024)
- XGBoost : A scalable tree boosting system (2016)
- LightGBM : A highly efficient gradient boosting decision tree (2017)
- CatBoost : Unbiased boosting with categorical features (2018)
- Long short-term memory (1997)
- Deep learning with long short-term memory networks for financial market predictions (2018)
- Attention is all you need (2017)
- Hierarchical multi-scale Gaussian transformer for stock movement prediction (2020)
- Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting (2019)
- Stock movement prediction from tweets and historical prices (2021)
- Stock selection via spatiotemporal hypergraph attention network (2021)
- Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns (2022)
- Qlib: An AI-oriented quantitative investment platform (2020) — link
- Open source cross-sectional asset pricing (2022)
- Advances in Financial Machine Learning (2018)
- The deflated S harpe ratio (2014)
- Reports of value’s death may be greatly exaggerated (2019)
- Survivorship bias and mutual fund performance (1996)
- The delisting bias in CRSP data (1997)
- A reality check for data snooping (2000)
- Stepwise multiple testing as formalized data snooping (2005)
- Price limit performance: Evidence from the Tokyo stock exchange (2013)
- Price limits and stock market volatility in China (2018)
- Price limit changes and market quality in the Chinese stock market (2005)
- Portfolio selection (1952)
- Active Portfolio Management (2000)
- Asset Management: A Systematic Approach to Factor Investing (2014)
- A well-conditioned estimator for large-dimensional covariance matrices (2004)
- Nonlinear shrinkage of the covariance matrix for portfolio selection (2017)
- Analytical nonlinear shrinkage of large-dimensional covariance matrices (2020)
- High dimensional covariance matrix estimation using a factor model (2008)
- Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? (2009)
- Machine learning and portfolio optimization (2018)
- Deep learning for portfolio optimization (2020)
- AlphaPortfolio: Direct construction through deep reinforcement learning and interpretable AI (2021)
- Deep portfolio optimization via distributional prediction of residual factors (2021)
- The pricing of options and corporate liabilities (1973)
- The jackknife and the bootstrap for general stationary observations (1989)
- The stationary bootstrap (1994)
- Resampling Methods for Dependent Data (2003)
- Modeling financial time-series with generative adversarial networks (2019)
- Quant GAN s: Deep generation of financial time series (2020)
- Time series simulation by conditional generative adversarial net (2019) — link
- Data augmentation approaches in natural language processing (2022)
- On the constrained time-series generation problem (2024)
- CVXPY : A P ython-embedded modeling language for convex optimization (2016)
- Conic optimization via operator splitting and homogeneous self-dual embedding (2016)
- Scikit-learn: Machine learning in P ython (2011)
- Adam: A method for stochastic optimization (2015)
- MOSEK Optimization Toolbox (2023) — link
- Types and Programming Languages (2002)
- Monads for functional programming (1995)
- A B ayesian approach to real estate assessment (1975)
- Stock price prediction via discovering multi-frequency trading patterns (2017)
- 60 years of portfolio optimization: Practical challenges and current trends (2014)
- Temporal relational ranking for stock prediction (2019)
0 of 76 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.