Crypto Stat-Arb

This repository studies market-neutral crypto statistical arbitrage with signed-graph clustering and walk-forward backtesting. It builds a residualized correlation graph after removing the market mode, clusters the graph with signed methods such as SPONGE and BNC, and trades cluster-level mean-reversion signals under explicit turnover and transaction-cost controls.

Repository layout

• stat_arb/: main research package for data loading, graph construction, clustering, signals, backtests, and reporting
• data/: processed market, volume, ETH, and correlation datasets used by the backtests
• pics/: diagnostic figures for clustering quality and exploratory analysis
• crypto_project.ipynb: exploratory notebook used during early research
• archived_research/: older exploratory artifacts retained for reference
• Crypto_Project_Report_Pre_Backtest.pdf: written report from the earlier research stage

Setup

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install numpy pandas scipy scikit-learn matplotlib statsmodels

Run the baseline SPONGE backtest:

python stat_arb/run_phase1.py

Run the clustering-method sweep:

python stat_arb/run_phase2.py

If you want to rerun the notebook cells that call CoinMarketCap, export your credential first:

export CMC_API_KEY=your_coinmarketcap_key

Methodology

The pipeline first aligns token prices, volumes, and ETH reference data, then builds a tradable universe subject to history and liquidity filters. Returns are residualized against the market mode with PCA, transformed into a signed k-nearest-neighbor correlation graph, and clustered with SPONGE, BNC, or signed spectral methods. Signals are generated from within-cluster mean reversion, normalized to target leverage, and evaluated in a walk-forward backtest with lagging, turnover controls, and transaction-cost assumptions to limit lookahead and overstatement.

Results

Primary outputs are written under stat_arb/reporting/ and include fold-level returns, turnover series, clustering sweep summaries, leaderboards, and the final report. The intended use is comparative research across clustering methods rather than a production-ready live trading engine.

Known limits

• Results are sensitive to crypto data quality, survivorship, and execution assumptions
• The checked-in notebook and archived artifacts reflect exploratory work and are less polished than the package backtest path
• Transaction costs and liquidity in crypto can change quickly enough to invalidate static assumptions

License

This project is distributed under the MIT License