quant-ml

A personal quantitative ML framework built across three Kaggle competitions as preparation for live Numerai competition.

Structure

quant-ml/
  core/               # Shared primitives — CV, eval metrics, book features, feature engineering
  competitions/
    optiver_volatility/   # Optiver Realized Volatility Prediction (2021) — order book, RMSPE
    optiver_close/        # Optiver Trading at the Close (2023) — closing auction, sequential buckets
    jane_street/          # Jane Street Real-Time Market Data Forecasting (2024) — streaming API
  numerai/            # Live Numerai Classic — era-based, Spearman IC, neutralization
  data/               # Local data only — gitignored, never committed

Competition Sequence

1. Optiver Volatility — Learn order book feature extraction, realized vol, RMSPE optimization
2. Optiver Close — Sequential bucket prediction, closing auction microstructure
3. Jane Street Real-Time — Streaming inference API, anonymized responders
4. Numerai — Live, sustained, era-based cross-sectional prediction

Core Modules

Module	Purpose
core/book.py	WAP, spread, imbalance, log returns, realized vol
core/cv.py	Purged K-Fold, walk-forward splits
core/eval.py	IC, ICIR, Sharpe, RMSPE, max drawdown
core/features.py	Z-score, rank normalize, neutralization, lag features

Setup

pip install -r requirements.txt

Data is not included. Download from Kaggle and place under data/<competition-name>/.