AI Engineering · Case study 10

WQ Auto Alphas — LLM-Driven Alpha Research

An LLM generating quant signals, grounded in a RAG knowledge bank of research papers.

The problem

Alpha research is a search problem: generate candidate expressions, simulate them, keep what survives. Doing that by hand is slow and biased toward whatever you thought of last.

The build

A miner that authenticates against the WorldQuant BRAIN API, uses Gemini to generate alpha expressions grounded in a SQLite RAG knowledge bank of alpha examples and research PDFs, runs simulations, and writes a JSONL audit log of every generated alpha and its result. Duplicate checks run against the knowledge bank before submission.

Architecture

— one teal highlight: the constraint filter

What I found

The generator has to know the platform's constraints, not just finance: it filters event-like fields from prompts and rejects time-series operators wrapped around event inputs, because BRAIN rejects them. Most of the engineering is teaching the LLM what won't compile.

The reps

Building the knowledge bank importer to ingest research PDFs and skip scripts, pickles and generated artifacts — retrieval quality is entirely a function of what you let in.

Results

  • RAG
    SQLite knowledge bank over papers + examples
  • JSONL
    full audit log of alphas and results
  • Auto
    generate → simulate → log loop

What's verified

Every case study carries this block

What was measured, and what was not. Nothing in the second column is estimated into the first.

Measured

The generation and simulation loop runs against the live BRAIN API with credentials supplied locally.

Not executed

No alpha performance figures are published here — platform results stay on the platform.

PythonGemini APIWorldQuant BRAIN APISQLite RAGOllama
View repository →