No. This repository is a research and engineering baseline. It does not connect to a broker, place orders, manage live risk, or provide trading recommendations.
No. The paper used proprietary market data that cannot be redistributed. The public repository therefore provides synthetic-data and public-data reproduction paths.
Not from the public repository alone, because the original proprietary data is not included. You can reproduce the pipeline structure and qualitative behavior with synthetic data, and you can add your own data through the loader interfaces.
Run:
make paper CONFIG=configs/small.yamlThen open notebooks/public_factor_ic.ipynb if you want a public-data walkthrough.
Financial panels have missing observations, suspensions, holidays, limit-up / limit-down cases, and varying security histories. Mask-aware tensors keep factor computation explicit and safer.
The legacy factor library provides a large practical testbed for tensorized factor computation. It also gives researchers many small signals to ablate, inspect, and replace.
Good contributions include:
- benchmark results
- setup feedback
- documentation fixes
- notebook cleanup
- tests for edge cases
- clearer examples
Use the Benchmark result issue template and include:
- commit SHA
- command
- Python version
- PyTorch version
- CPU/GPU model
- CUDA availability
- printed result table
Public-data examples are mainly for workflow demonstration and reproducibility. They may have survivorship bias, missing fields, adjusted-price differences, stale liquidity assumptions, and transaction-cost limitations. Treat them as examples, not claims of deployable alpha.
Yes. A good adapter should produce a Panel with consistent open, high, low, close,
volume, vwap, and mask fields, and should document missing fields and adjustments.