Validation · #357857 · Crypto

Cointegration stat-arb

Dynamic Cointegration-Based Pairs Trading Strategy in the Cryptocurrency Market, 2021

Masood Tadi & Irina Kortchmeski · arXiv q-fin

The rule

Form a Johansen-cointegrated basket of BTC-denominated alt-coins; trade its Z-score — enter when |Z| > 2, exit when |Z| < 1.

Fails 7.94→−5.90 claimed → measured Sharpe

7.94

Claimed Sharpewith maker rebates

−5.90

Net Sharperealistic 7.5 bps taker

17.70

Replicated gross Sharpein-sample minute-mining

AssetCrypto

StrategyJohansen basket

Universe7 alts / BTC

Period2018–2024 OOS

Costs7.5 bps taker

Assets tradedBTC ETH EOS LTC XRP TRX ADABinance USDT pairs, each alt re-denominated in BTC. BCH (the paper's 8th coin) dropped — the Nov-2018 fork breaks its ticker history.

The exact rules

Formation window	→	Fit a Johansen-cointegrated basket of the BTC-denominated alts; track its spread Z-score
\|Z\| > 2 (spread stretched)	→	Enter — trade the basket back toward its mean
\|Z\| < 1 (spread reverted)	→	Exit to flat

The backtest, re-run

Inside the model

The validation ladder

C0	Replicate	Gross 17.70 — reproduces the in-sample minute-mining artifact (claimed 7.94).
C1	Honesty	Realistic 7.5 bps taker → −5.90; the strategy turns over ~6,000×/yr.
C2	Deflate	The gross edge is in-sample-mined over a basket on 131k minute bars; honest net is −5.90.
C5	Frictions	Cost sweep 0bp → 8.0, 7.5bp → 1.9, 15bp → −5.0, 30bp → −17.6. Breakeven ≈ maker-rebate territory.
C6	Decay	Frozen basket gross 7.6 (2019) → 1.0 (2020–22) → negative (2023). Cointegration breaks OOS.
C7	Originality	BTC-neutral (β ≈ −0.02, R² 1%) — genuine relative-value, but α t = 1.8 (not robust) and decays.

Reproduces (gross)Real after costsHolds out-of-sampleCrisis-robustOriginal vs BTC

How we rebuilt it

Data: Binance USDT minute, re-denominated in BTC (alt/BTC) per the paper — the step that makes the alts cointegrate.
Method: Johansen-basket cointegration → OU half-life Z-window → event-driven |Z| backtest.
Universe: BitMEX 8 → Binance 7 (BCH dropped — Nov-2018 fork breaks its ticker).
Regime stated: None — minute window 2018–2019; relationship persistence tested hourly to 2024.

Deviations from the paper

Data: BitMEX (XBT-denominated) → Binance USDT, BTC-denominated; same window 2018-09→2019-10.
Paper assumed BitMEX maker rebates (negative fees) → realistic 7.5 bps taker (the key honest correction).
Sharpe convention undisclosed → daily returns ×√365 (crypto trades 24/7).

ProvenanceBeta

Engine: v1
Blocks: 8 blocks · 7 new (cointegration · pairs · crypto adapter), 1 reused (metrics)
Data: binance/7coin_btc_denom · minute 2018–19, hourly to 2024
Source: Binance market data ↗ · via ccxt ↗
Tests: 14/14 passing · statistical + adversarial-code QA clean