Looking for feedback on ml-quant-trading, an open-source PyTorch stack for ML factor research

[initial-d]

🏷️ Discussion Type

Question

Body

Hi GitHub Community,

I recently open-sourced ml-quant-trading, a research-oriented PyTorch stack for ML-enhanced multi-factor quantitative trading experiments:

https://github.com/initial-d/ml-quant-trading

The goal is not to claim live trading profitability. It is meant to be a reproducible engineering baseline for factor research, public-data experiments, and backtesting pipeline design.

What is included:

• 213 factor dimensions
• mask-aware tensor factor primitives
• limit-up / limit-down / halt bias correction
• MLP and Transformer baselines
• Markowitz portfolio construction
• vectorized backtesting and metrics
• synthetic and public-data demos
• CI, tests, benchmark scripts, and contribution templates

Paper:
https://arxiv.org/abs/2507.07107

I would especially appreciate feedback from people who care about research software, reproducible notebooks, or performance benchmarking:

• Does the first-run setup make sense?
• Are the docs clear enough for someone new to this domain?
• Would anyone be willing to share CPU/GPU benchmark results?
• Are there public-data reproduction paths you would like to see added?
• Are there repository structure or CI improvements that would make this more contributor-friendly?

I searched for similar discussions before posting and did not find an existing thread for this project. Thanks for any feedback or criticism.

[initial-d]

Quick update: v0.1.0 is now published as the first public research baseline:

https://github.com/initial-d/ml-quant-trading/releases/tag/v0.1.0

I also opened a dedicated thread for benchmark results, public-data reproductions, and setup feedback:

initial-d/ml-quant-trading#13

Most useful feedback right now: CPU/GPU benchmark numbers, public-data reproduction notes, mask/limit-up/halt edge cases, and docs/setup issues.

[initial-d]

Update: the repo now has a public-data mini reproduction note in addition to the benchmark board:

https://github.com/initial-d/ml-quant-trading/blob/main/docs/public_data_mini_reproduction.md

It uses yfinance, 10 liquid US equities, a six-factor subset, and one-day forward rank IC. The goal is reproducibility/API validation, not profitability claims. More useful next feedback: larger public universes, ETF examples, A-share public-provider runs, CUDA/Linux benchmarks, and setup friction.

[initial-d]

Thanks, Vishal — this is exactly the kind of practical feedback I was hoping for.

I adopted the main suggestions in this commit: initial-d/ml-quant-trading@ac19775

What changed:

• expanded Start Here with clone-to-demo commands, expected runtime/hardware, expected artifact directories, typical successful output, and warning guidance
• added a VS Code / Codespaces Dev Container for reproducible setup
• added and linked a Mermaid architecture diagram for data -> factors -> model training -> portfolio construction -> backtest -> reports
• expanded the benchmark board/template to include CUDA version alongside CPU/GPU/PyTorch/OS
• added expected-output checks to the public-data mini reproduction note, including ticker universe, date range, factor subset, expected IC table, and synthetic fallback seed
• refreshed the roadmap and newcomer entry points around benchmark reports and public-data reproductions

The CI run is green across Python 3.9, 3.10, and 3.11 for lint, pytest, and the CLI smoke test: https://github.com/initial-d/ml-quant-trading/actions/runs/28732179750

Really appreciate the review — it turned into a much better first-contributor path.

[initial-d]

Thank you, Vishal — I really appreciate the encouragement.

That was exactly the goal: make the repository easier to evaluate in the first 10 minutes, and easier to contribute to after that. The next things I want to collect are real benchmark reports from different machines and a few more public-data reproductions, so the project can grow from a maintainer baseline into a community-comparable research reference.

Thanks again for taking the time to review it so concretely — your checklist made the next iteration much sharper.

[swapnil5053]

The tensor factor primitives and the bias correction setup seem very interesting. I have been going through the repo and the paper, and it looks like a strong baseline.

To address a few of your questions:

Setup and Documentation: The setup is logical, and the documentation is quite clear. Including a Dockerfile would be helpful to avoid random local dependency issues.

Benchmarks: I worked on Fiber some time ago, which showed me how difficult it can be to get reliable performance metrics. I appreciate good profiling because of that experience. I would be happy to run your benchmark scripts over the weekend and report my CPU and GPU results.

Structure and Continuous Integration: Setting up something like CodeRabbit for automated pull request reviews could be a valuable addition. It reduces the friction in reviewing community contributions. Also, using Jira or a GitHub Project board to track upcoming features and bugs would help outside contributors see where they can take on tasks.

[initial-d]

Thanks, Swapnil — this is very helpful feedback, and I appreciate you taking a look through both the repo and the paper.

I adopted the Docker part right away in this commit: initial-d/ml-quant-trading@d110ee8

What changed:

• added a root Dockerfile for a reproducible CPU dev/test environment
• added .dockerignore
• updated the VS Code / Codespaces Dev Container to reuse the same Dockerfile
• added Docker commands for tests, lint, the synthetic pipeline, benchmarks, and public-data validation: https://github.com/initial-d/ml-quant-trading/blob/main/docs/docker.md
• linked the Docker path from README and CONTRIBUTING

CI is green across Python 3.9, 3.10, and 3.11 for the commit: https://github.com/initial-d/ml-quant-trading/actions/runs/28759261007

On the benchmark side, CPU/GPU results would be extremely useful. The best place to share them is the benchmark/reproduction discussion:

initial-d/ml-quant-trading#13

or the benchmark collection issue:

initial-d/ml-quant-trading#7

For CodeRabbit / automated PR review and a project board, I opened a tracking issue rather than connecting a third-party service immediately: initial-d/ml-quant-trading#17

I want to evaluate those after the first external PRs arrive, so the repo gets contributor support without adding too much review noise or service dependency.

Thanks again — this turned into a concrete reproducibility improvement.

[v-ember]

Hi @initial-d thanks for sharing with the community!

Unfortunately, we currently do not allow self-promotion, advertising, or solicitation in Community Discussions. We want to make sure there is space for users to ask questions without overwhelming them with other conversations. Thank you for helping us maintain a productive and tidy community for all our members.