A cost-aware cross-sectional factor research stack built around reproducible reports

[initial-d]

I maintain ml-quant-trading, a PyTorch research stack for cross-sectional factor modeling and portfolio backtesting.

The project uses panel weights rather than single-instrument order simulation, so it has a small vectorized engine designed around:

• 213 factor dimensions and mask-aware tensor operations
• limit-up/down, halt, and missing-data handling
• MLP and Transformer baselines
• Markowitz portfolio construction
• explicit turnover, costs, slippage, drawdown, and baseline comparisons
• synthetic, yfinance, Baostock, and AkShare reproduction paths
• machine-readable Markdown/JSON validation reports

The latest public CSI 300 report is here:
https://github.com/initial-d/ml-quant-trading/blob/main/docs/validation_dashboard.md

One useful result is deliberately unglamorous: a naive daily factor portfolio had positive gross evidence but excessive turnover consumed much of the edge. A transparent holding-buffer rule reduced turnover from 0.3627 to 0.1397 and improved the 7 bps net result. Failed models and data limitations remain visible.

There is also a zero-account Colab:
https://colab.research.google.com/github/initial-d/ml-quant-trading/blob/main/notebooks/quickstart_colab.ipynb

I am sharing this here because the backtesting.py community tends to care about semantics and reproducibility, not just equity curves. Critique of cost accounting, weight timing, public-data assumptions, and report design is very welcome. This is research software, not an investment claim.