Open-source PyTorch multi-factor stack with public CSI 300 validation

[initial-d]

I maintain ml-quant-trading, an open-source PyTorch research stack for cross-sectional multi-factor workflows.

It is complementary to the reinforcement-learning path represented by FinRL: the repository focuses on a factor → supervised model → portfolio optimizer → vectorized backtest pipeline, with 213 factor dimensions, mask-aware tensor operations, MLP/Transformer baselines, Markowitz construction, transaction costs, and public-data reports.

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

The latest maintained public-data run uses current CSI 300 constituents and AkShare daily data from 2021-01-04 to 2024-12-31. At 7 bps effective cost, a buffered factor portfolio reduced turnover from 0.3627 to 0.1397 and returned 22.20% annualized versus 17.75% for the equal-weight baseline. The naive factor portfolio lost much of its gross edge to turnover.

Full result and caveats:
https://github.com/initial-d/ml-quant-trading/blob/main/docs/validation_dashboard.md

Important limitations: this is current-constituent public data rather than point-in-time membership; it is not an exact paper reproduction, live-trading evidence, or investment advice. Negative model results remain in the report.

I would be interested in a clean comparison between a FinRL policy and these cross-sectional baselines on the same public panel, cost grid, and walk-forward split. The repository exports Markdown/JSON reports and welcomes both successful and failed reproductions.