Crypto Statistical Arbitrage (4H Binance Spot): Momentum vs. Short-Horizon Reversal

Research backtest comparing cross-sectional momentum vs short-horizon reversal (mean reversion) on a liquid Binance USDT spot universe using 4-hour candles, with turnover-based transaction costs.

• Momentum: winners keep winning (trend continuation)
• Reversal: losers bounce (mean reversion)

This is a research backtest for learning and evaluation — not investment advice.

---

Method (high level)

• Universe: liquid USDT spot pairs (filtered)
• Signals: momentum (lookback L) vs reversal (lookback H)
• Portfolio: cross-sectional long/short (top 20% vs bottom 20%)
• Execution: daily rebalancing, 1-bar lag (no look-ahead)
• Costs: transaction costs modeled from turnover (baseline 20 bps)

---

Results (snapshot from latest run)

Momentum sweep (L ∈ {6, 12, 24, 42, 84, 168})

Momentum was negative across all tested horizons in this sample.

Reversal sweep (H ∈ {1, 2, 3, 6})

Reversal was strongly positive, with H=3 (~12 hours) best in the unfiltered run.

---

Cost sensitivity (unfiltered reversal, H=3)

H	cost (bps)	Sharpe	CAGR	MaxDD	AvgTurnover	FinalEquity
3	20	3.596	2.062	-0.261	0.255	28.704×
3	40	0.159	-0.004	-0.728	0.255	0.989×
3	60	-2.910	-0.677	-0.974	0.255	0.034×

Takeaway: performance is strong at 20 bps but drops sharply at higher costs → execution matters.

---

Diagnostic: why reversal appears real (bucket test)

Each bar, assets are ranked by reversal signal and split into 5 buckets (0=lowest, 4=highest).
If mean reversion exists, the most “oversold” bucket should bounce more next bar.

---

Train/Test split (cutoff 2024-01-01)

---

Execution improvement: liquidity filter

Optional filter keeps only assets in the top 50% of volume rank each bar (reduces trading in thin / hard-to-trade names).

---

Equity / Turnover / Cost plots

Unfiltered (H=3):

• 
• 
• 

Liquidity-filtered:

• 
• 
• 

---

Reproducibility note (why results can change “day to day”)

If you re-download OHLCV or re-filter the universe (coverage, top-volume list, etc.), the asset set and sample period can shift slightly — changing the best H/L and performance.

To lock results:

1. Freeze a start/end date
2. Save the universe tickers used
3. Cache OHLCV locally (parquet/csv) and reuse it

---

Project structure

.
├── notebooks/
│   └── Crypto_Statistical_Arbitrage.ipynb
├── src/
│   └── crypto_statarb.py
├── figures/
│   └── (pngs used in README / WordPress)
├── requirements.txt
└── .gitignore


## Quick start
pip install -r requirements.txt  
Open `notebooks/Crypto_Statistical_Arbitrage.ipynb` in Colab/Jupyter.

## License
MIT (see `LICENSE`).