WorldQuant Alpha Research Toolkit
An engineered workflow for WorldQuant BRAIN Alpha research: candidate generation, multi-metric evaluation, and human-approved submissions.
- Candidate generation
- Syntax and structure checks
- BRAIN simulation
- Multi-metric evaluation
- Bounded tuning
- Report
- Human-approved submission
Why I built it
A good-looking single backtest on WorldQuant BRAIN does not prove that an Alpha is usable, and testing expressions one by one makes it hard to explain why a candidate passed or was dropped. I wanted to turn the research process into a traceable pipeline.
Design
- One workflow from start to finish: A Python CLI connects candidate generation, local syntax and structure checks, BRAIN simulation, metric evaluation, bounded tuning, and reporting.
- Consistent multi-metric evaluation: Sharpe, Fitness, turnover, drawdown, and autocorrelation are assessed against a shared set of criteria. Candidates that pass the baseline metrics can then go through bounded robustness probes.
- Recoverable batches: Research batches are persisted and support pre-run checks and backup recovery, along with state inspection, resume, and cancellation through LangGraph.
- People decide what gets submitted: The tool produces reports and recommendations. A person always confirms whether to submit.
My role
Organized the research steps into recoverable batches; defined boundaries for local checks, platform simulations, multi-metric evaluation, and bounded tuning; and kept human approval and report lookup in the workflow.
How it was validated
- Candidates pass local syntax and structure checks before they can enter platform simulation.
- Decisions do not rely on a single in-sample backtest: bounded robustness probes follow multi-metric evaluation.
- Each batch keeps a report so the reasons for passing or dropping a candidate can be reviewed.
What is not public
Expressions, data fields, and submission details are not disclosed. This page covers only the methods and engineering practices.