LavaQuant

Alpha research lab. Write WorldQuant-style expressions, backtest them, compare against BRAIN and Numerai. Next.js + FastAPI + LightGBM.

Unaffiliated with WorldQuant.

Live: quant.drew.fun

The result on first load is simulated sample data. Run a backtest for real numbers.

What it does

Type an alpha like group_neutralize(rank(-returns), sector), hit run, get IC, Sharpe, drawdown and a few charts in about 10 seconds.

• Editor with highlighting, autocomplete and live error checking
• Backtests on an S&P 500 sample, sector ETFs, forex and commodities
• Optional submit to WorldQuant BRAIN to compare scores
• LightGBM / ridge models with time-series CV
• Cmd+K palette, and shareable links (?alpha=...) that open and run an alpha

Metrics

Full definitions are in the app under Docs.

IC	Daily rank correlation between the alpha and forward returns
IC-IR	mean(IC) / std(IC). Daily, not annualized
Sharpe	mean / std of daily return, times √252
Turnover	Share of the book traded per day
Fitness (est.)	Sharpe × √(|return| / max(turnover, 0.125)), WorldQuant's formula on local numbers
Score	0-100 blend of the above. My own heuristic, not an industry metric

Worth knowing:

• Returns are gross of costs.
• The "platform fit" thresholds are rules of thumb, not official Numerai or WorldQuant criteria.
• Local data is a ticker sample from Yahoo. It won't match BRAIN's numbers, which is why the BRAIN button exists.

How it's built

Next.js on Vercel calls a FastAPI backend on Railway. The DSL is parsed with ast against a function whitelist, no eval. BRAIN logins pass through the backend per request and aren't stored or logged.

Run locally

Python 3.9+ and Node 18+.

cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload

cd frontend
npm install
npm run dev

Frontend is on :3000, backend on :8000.

Deploy

Vercel with root frontend, Railway with root backend (Dockerfile included).

Variable	Where
NEXT_PUBLIC_API_URL	Frontend	Backend URL, e.g. https://x.up.railway.app
NEXT_PUBLIC_SITE_URL	Frontend	Public URL, for link previews
ALLOWED_ORIGINS	Backend	Comma-separated CORS origins. Has defaults
CACHE_DIR	Backend	Price data cache. Optional

NEXT_PUBLIC_* is baked in at build time, so redeploy after changing it.

Limitations

• Prices come live from Yahoo and the cache is wiped on deploy, so the first request after one is slow.
• No costs, slippage or borrow fees.
• ML CV only trains on data before each validation fold, with an embargo gap. Not a full purged k-fold.

Research only, not investment advice.