mlquant

A reproducible PyTorch research stack for machine-learning multi-factor trading: 213 factors, bias correction, portfolio optimization, vectorized backtesting, and System-One-style portfolio decisions.

Install and run

python -m pip install mlquantx
mlquant demo

The demo needs no market-data account or API key. It runs the deterministic synthetic pipeline from data generation through 213 factor dimensions, model training, portfolio construction, cost-aware backtesting, and Markdown/JSON report generation. The default config ships inside the wheel, so the command works outside a repository checkout.

Customize or contribute

The wheel is the fastest way to try the project. Clone the repository when you want to change factors, models, portfolio constraints, data sources, or backtest assumptions:

git clone https://github.com/initial-d/ml-quant-trading.git
cd ml-quant-trading
python -m pip install -e '.[dev]'

If the demo saves you setup time, consider starring the repository or sharing a reproducible run.

What is included

• 204 hand-crafted factors plus 9 curated Alpha101-style factors
• mask-aware PyTorch tensor primitives for cross-sectional panels
• limit-up, limit-down, halt, and missing-data bias handling
• MLP and Transformer research baselines
• constrained Markowitz portfolio construction
• vectorized backtesting with turnover and transaction costs
• AkShare, Baostock, yfinance, and deterministic synthetic data paths
• auditable public-data validation reports, including negative results
• mlquant systemone decide for typed portfolio-intent decisions from structured research state

Start here

• Source and full documentation
• Google Colab quick start
• Public validation dashboard
• System-One-compatible portfolio decisions
• Research card and limitations
• 100,000-row synthetic dataset
• 213-input MLP checkpoint
• Paper: arXiv:2507.07107

Research boundary

mlquant is research and educational software. It is not investment advice or a production trading system. Synthetic smoke tests verify engineering behavior, not profitability. Public-data backtests depend on data quality, survivorship, transaction costs, slippage, and modeling assumptions and do not represent live or guaranteed out-of-sample performance.

License

MIT. See the repository license.