Agent Quant Benchmark Challenge: can your coding agent reproduce this pipeline?

[initial-d]

I turned ml-quant-trading into an Agent Quant Benchmark Challenge.

Question:

Can your coding agent reproduce a 213-factor, cost-aware quant research pipeline, preserve the evidence bundle, and avoid calling a backtest "alpha"?

This is not a trading-signal contest and not investment advice. The challenge is about reproducibility discipline on a real quant research codebase:

• zero-account synthetic smoke test: mlquant demo
• fixed protocol v1 CPU benchmark
• DeepSeek Harness benchmark path
• public-data validation with costs, turnover, and caveats
• redacted private evaluation notes for users who cannot disclose proprietary data

Start here:

• Challenge page: https://github.com/initial-d/ml-quant-trading/blob/main/docs/agent_quant_benchmark_challenge.md
• Colab quickstart: https://colab.research.google.com/github/initial-d/ml-quant-trading/blob/main/notebooks/quickstart_colab.ipynb
• DeepSeek Harness recipe: https://github.com/initial-d/ml-quant-trading/blob/main/docs/deepseek_harness_recipe.md
• Private evaluation checklist: https://github.com/initial-d/ml-quant-trading/blob/main/docs/private_evaluation_checklist.md

Useful reports can be successful or failed. Please keep the evidence boundary clear: command, commit SHA, environment, generated artifacts, warnings, costs, data source category, and caveats.

Do not post credentials, account identifiers, proprietary datasets, live positions, order logs, private alpha definitions, internal hostnames, or production strategy rules.

If you run it with a coding agent or quant-agent workflow, feel free to post the report here or use the structured issue templates linked from the challenge page.

[initial-d]

Quick participation path:

1. Pick one track from the challenge page.
2. Run it with a human workflow, coding agent, or DeepSeek Harness.
3. Post a report with commit SHA, exact command, environment, generated artifacts, warnings, and caveats.

Copy/paste prompt for agents:

Read AGENTS.md, README.md, docs/reality_check.md, docs/benchmarking.md,
docs/agent_reproducibility.md, and docs/agent_quant_benchmark_challenge.md.
Run one challenge track exactly as documented. Preserve the commit SHA, command,
environment, generated Markdown/JSON artifacts, warnings, retries, and caveats.
Do not describe runtime, synthetic data, or public-data backtests as trading
alpha. Draft an issue-ready report using the appropriate template.

Failed runs are welcome if they are specific. Please redact credentials, account identifiers, proprietary datasets, live positions, order logs, private alpha definitions, internal hostnames, and production strategy rules.