agiprolabs/cointegration-analysis

Cointegration Analysis

Cointegration testing identifies pairs of assets that share a long-run equilibrium relationship, enabling statistical arbitrage and pairs trading strategies.

What Is Cointegration?

Two price series are cointegrated when they are individually non-stationary (random walks) but a linear combination of them is stationary (mean-reverting). Intuitively, the prices may wander apart temporarily but are pulled back to an equilibrium spread over time.

Cointegration vs Correlation

| Property | Correlation | Cointegration | |---|---|---| | Measures | Short-term co-movement | Long-run equilibrium | | Stationarity | Requires stationary returns | Works with non-stationary prices | | Time horizon | Can change rapidly | Stable over months/years | | Trading use | Momentum/trend signals | Mean-reversion pairs trades | | Failure mode | Breaks in regime changes | Breaks on structural shifts |

Two assets can be highly correlated but not cointegrated (e.g., two unrelated uptrends). Conversely, cointegrated assets may have low short-term correlation during temporary divergences — which is exactly when pairs trades are entered.

Why It Matters

• Pairs trading: Long the underperformer, short the outperformer, profit on convergence
• Statistical arbitrage: Systematic mean-reversion on spread z-scores
• Spread trading: Trade the spread directly as a synthetic instrument
• Risk hedging: Cointegrated hedge ratios minimize tracking error over time

Methods

1. Engle-Granger Two-Step

The most common approach for two series.

Step 1 — Regress Y on X using OLS:

Y_t = α + β * X_t + ε_t

Step 2 — Test the residuals ε_t for stationarity using the ADF test.

• If residuals are stationary (p < 0.05) → Y and X are cointegrated
• β is the hedge ratio for the pairs trade
• α is the long-run mean of the spread

Important: Engle-Granger critical values differ from standard ADF critical values. For n=2 series: 1% = -3.90, 5% = -3.34, 10% = -3.04.

Asymmetry warning: Testing Y~X can give a different result than X~Y. Always test both directions and use the stronger result.

from scipy import stats
import numpy as np
from statsmodels.tsa.stattools import adfuller

# Step 1: OLS regression
slope, intercept, _, _, _ = stats.linregress(x_prices, y_prices)
hedge_ratio = slope

# Step 2: Test residuals
residuals = y_prices - hedge_ratio * x_prices - intercept
adf_stat, p_value, _, _, crit_values, _ = adfuller(residuals, maxlag=None, autolag="AIC")

cointegrated = p_value < 0.05

2. Johansen Test

Tests multiple series simultaneously and returns the number of cointegrating relationships. More powerful than Engle-Granger for >2 series.

• Based on a VAR model: ΔY_t = Π·Y_{t-1} + Σ Γ_i·ΔY_{t-i} + ε_t
• Tests the rank of the Π matrix
• Uses trace test and maximum eigenvalue test
• Returns: number of cointegrating vectors and the vectors themselves

from statsmodels.tsa.vector_ar.vecm import coint_johansen

# data: T×N array of price series
result = coint_johansen(data, det_order=0, k_ar_diff=1)

# Trace statistic vs critical values (90%, 95%, 99%)
trace_stats = result.lr1          # Trace statistics
trace_crit = result.cvt           # Critical values
max_eigen_stats = result.lr2      # Max eigenvalue statistics
max_eigen_crit = result.cvm       # Critical values

# Cointegrating vectors
coint_vectors = result.evec

3. Phillips-Ouliaris

Similar to Engle-Granger but uses Phillips-Perron style test statistics instead of ADF. More robust to heteroskedasticity and serial correlation in the residuals. Available via statsmodels.tsa.stattools.coint.

from statsmodels.tsa.stattools import coint

# Returns: test statistic, p-value, critical values
t_stat, p_value, crit_values = coint(y_prices, x_prices)
cointegrated = p_value < 0.05

Practical Workflow

Step 1: Screen Pairs by Correlation

Pre-filter using Pearson correlation > 0.7 to reduce the number of cointegration tests (which are more expensive).

Step 2: Test Cointegration

Run Engle-Granger in both directions. Use p < 0.05 threshold.

Step 3: Estimate Hedge Ratio

Use OLS for simplicity. For production, consider Total Least Squares or Dynamic OLS (see references/methodology.md).

Step 4: Compute Spread

spread = y_prices - hedge_ratio * x_prices - intercept
z_score = (spread - spread.mean()) / spread.std()

Step 5: Test Spread for Mean Reversion

• ADF test: p < 0.05 confirms stationarity
• Hurst exponent: H < 0.5 indicates mean reversion (H ≈ 0.5 = random walk)
• Half-life: λ from AR(1) on spread; half-life = -ln(2)/ln(λ)
  • Viable pairs: half-life between 5 and 60 days

Step 6: Trade the Spread

If the spread is mean-reverting, it is a viable pairs trade candidate. See references/pairs_trading.md for entry/exit rules and risk management.

Rolling Cointegration

Cointegration relationships can break down over time due to structural changes, regime shifts, or evolving market dynamics.

Rolling Window Approach

Test cointegration on rolling 60–90 day windows:

window = 60
rolling_pvalues = []
rolling_hedges = []

for i in range(window, len(prices)):
    y_win = y_prices[i - window:i]
    x_win = x_prices[i - window:i]
    _, p_val, _ = coint(y_win, x_win)
    slope, intercept, _, _, _ = stats.linregress(x_win, y_win)
    rolling_pvalues.append(p_val)
    rolling_hedges.append(slope)

Monitoring Signals

| Signal | Healthy | Warning | Stop Trading | |---|---|---|---| | Rolling p-value | < 0.05 | 0.05–0.10 | > 0.10 | | Hedge ratio drift | < 10% change | 10–25% change | > 25% change | | Spread half-life | 5–60 days | 60–120 days | > 120 days or < 5 |

Crypto Pairs Candidates

Layer-1 Correlation

• SOL vs ETH — L1 sector beta, often cointegrated during trending markets
• SOL vs AVAX — alternative L1 correlation

Stablecoins

• USDC vs USDT — should be perfectly cointegrated (peg arbitrage)
• Useful as a sanity check for your cointegration pipeline

Liquid Staking Derivatives

• mSOL vs jitoSOL — both track SOL staking yield
• stSOL vs mSOL — Lido vs Marinade staking

Same-Sector Tokens

• DEX tokens: RAY vs ORCA
• Lending tokens: cross-protocol comparison
• Meme tokens: rarely cointegrated, high risk

Common Pitfalls

1. Spurious cointegration — Two trending series (both up in a bull market) may appear cointegrated. Always test on sufficient data (>200 observations) and check out-of-sample stability.

2. Structural breaks — A fundamental change (protocol upgrade, tokenomics change) can permanently break cointegration. Monitor rolling p-values.

3. Look-ahead bias — Estimating the hedge ratio on the full sample and then backtesting on the same sample inflates results. Always use walk-forward estimation.

4. Too-short sample — Cointegration tests need >100 observations minimum, ideally >200, to have reasonable power.

5. Ignoring transaction costs — Pairs trades involve 4 transactions per round trip. At 0.3% per leg, that is 1.2% in costs that the spread must overcome.

6. Asymmetric cointegration — The relationship may only hold in one direction or one regime. Consider threshold cointegration models for production use.

Integration with Other Skills

• correlation-analysis — Pre-screening pairs by correlation before cointegration testing
• mean-reversion — Trading the cointegrated spread using mean-reversion entry/exit rules
• vectorbt — Backtesting pairs strategies with walk-forward validation
• regime-detection — Identifying when cointegration regimes shift
• volatility-modeling — Spread volatility forecasting for dynamic position sizing

Files

References

• references/methodology.md — Engle-Granger details, Johansen derivation, hedge ratio estimation methods, spread construction
• references/pairs_trading.md — Entry/exit rules, risk management, performance metrics, crypto-specific considerations

Scripts

• scripts/test_cointegration.py — Full cointegration test pipeline with ADF, Hurst, half-life, rolling stability, and demo mode
• scripts/pairs_backtest.py — Walk-forward pairs trading backtest with synthetic data and performance reporting