Cointegration-Based Pairs Trading: Still Viable in 2024?

1. The Rise and Decline of Pairs Trading

Pairs trading — buying the undervalued stock in a cointegrated pair while shorting the overvalued one — was famously systematised by Nunzio Tartaglia's quantitative group at Morgan Stanley in the 1980s. Gatev, Goetzmann, and Rouwenhorst (2006) published the seminal academic study documenting robust profitability from 1962 to 2002. The strategy's appeal was straightforward: it was market-neutral, required no directional view, and was grounded in a rigorous statistical framework (Engle and Granger's cointegration theory).

Since publication, however, the evidence has been mixed. Krauss (2017) surveyed the literature and found that profitability had declined in US equities from the mid-2000s onward. Do and Faff (2010, 2012) documented a similar pattern and attributed it to increasing market efficiency: as more participants adopted the strategy, the mispricings it exploited were arbitraged away faster. Rad, Low, and Faff (2016) compared distance, cointegration, and copula methods and found that cointegration-based approaches had higher Sharpe ratios historically but were also more sensitive to the estimation period.

We revisit this question with an additional decade of data, covering the period 2005–2024, which includes the GFC, the post-crisis low-volatility era, the COVID crash, and the 2022 rate-hiking regime. Our focus is on diagnosing why cointegration-based pairs trading has declined and whether adaptive methods can restore profitability.

2. Methodology

We follow the standard two-phase protocol. In the formation period (12 months), We screen all pairs within the S&P 500 for cointegration using the Engle-Granger two-step method. For each pair, We regress the log price of stock A on the log price of stock B, then test the residual for stationarity using the ADF test at the 5% significance level. Pairs that pass are ranked by the speed of mean reversion (estimated from the ADF coefficient), and the top 20 pairs are selected for trading.

In the trading period (6 months), We compute the z-score of the spread for each pair using the formation-period parameters. We enter when |z| > 2.0, exit when z crosses zero, and stop-loss when |z| > 4.0 or after 50 trading days without convergence. Position sizes are equal-dollar long-short. Transaction costs are set at 10 basis points per round trip.

from statsmodels.tsa.stattools import adfuller, coint
import numpy as np

def find_cointegrated_pairs(prices_df, significance=0.05):
    """
    Screen all pairs in a price DataFrame for cointegration.

    Args:
        prices_df: DataFrame with log prices, columns = tickers
        significance: ADF p-value threshold
    Returns:
        list of (ticker_a, ticker_b, adf_stat, half_life)
    """
    tickers = prices_df.columns
    n = len(tickers)
    pairs = []

    for i in range(n):
        for j in range(i+1, n):
            # Engle-Granger cointegration test
            score, pvalue, _ = coint(prices_df[tickers[i]],
                                      prices_df[tickers[j]])
            if pvalue < significance:
                # Estimate half-life from spread
                spread = (prices_df[tickers[i]] -
                         score * prices_df[tickers[j]])
                spread_lag = spread.shift(1).dropna()
                spread_diff = spread.diff().dropna()
                # AR(1) regression
                lam = np.polyfit(spread_lag.values,
                                spread_diff.values, 1)[0]
                half_life = -np.log(2) / lam if lam < 0 else np.inf

                if 5 < half_life < 60:  # tradeable range
                    pairs.append((tickers[i], tickers[j],
                                 pvalue, half_life))

    # Sort by half-life (faster reversion = better)
    return sorted(pairs, key=lambda x: x[3])

3. Performance Over Time

Table 1: Performance of the classic Engle-Granger pairs trading strategy by sub-period. "Convergence Rate" is the percentage of trades where the spread reverted to zero before the stop-loss or time limit was hit.

The decline is monotonic and substantial. The annualised Sharpe ratio has fallen from 1.24 in the earliest period to 0.19 in the most recent — a level indistinguishable from noise for a two-year evaluation window. Two diagnostics stand out: the number of cointegrated pairs found per formation period has declined from 87 to 31, and the convergence rate has fallen from 74% to 41%.

4. Why Cointegration Is Breaking Down

4.1 Structural Breaks

The primary cause of non-convergence is structural breaks in the cointegration relationship. A pair that was cointegrated during the formation period can lose cointegration during the trading period due to firm-specific events (earnings surprises, M&A activity, management changes), sector rotations, or macroeconomic regime shifts. When the cointegration relationship breaks, the spread does not mean-revert — it trends, and the pairs trade accumulates losses until stopped out.

We estimate the frequency of structural breaks using the Bai-Perron test applied to the spread residuals. In the 2005–2007 period, approximately 15% of traded pairs experienced a structural break during the 6-month trading window. By 2023–2024, this figure had risen to 38%. The increase is consistent with higher idiosyncratic volatility in US equities and the greater frequency of sector-disrupting events (AI adoption, pandemic effects, supply chain restructuring).

4.2 Crowding

The second cause is crowding: more participants running similar strategies extract the same mispricings, reducing the amplitude of the spread deviations and accelerating convergence for the trades that do work, but also reducing the overall opportunity set. The decline in the number of qualifying pairs — from 87 to 31 per formation period — is consistent with crowding compressing spreads below the threshold needed to cover transaction costs.

4.3 Factor Concentration

Many "cointegrated" pairs in US equities are actually pairs that share common factor exposures (size, value, momentum, sector). When the dominant factors rotate rapidly — as occurred in 2020–2022 with the growth-to-value rotation — pairs that appeared cointegrated due to shared factor exposure can diverge sharply. This is not a true cointegration breakdown but a failure of the screening process to identify genuinely fundamental economic relationships versus statistical artefacts of common factor loading.

5. Adaptive Methods

We test three adaptations designed to address these failure modes:

Rolling re-estimation. Rather than fixing parameters for the entire 6-month trading window, We re-estimate the cointegrating vector monthly using a rolling 252-day window. This allows the model to track slow parameter drift but does not help with sudden structural breaks.

Structural break detection. We implement a real-time CUSUM test on the spread residuals. When the CUSUM statistic exceeds a critical threshold, We classify the pair as "broken" and close the position immediately rather than waiting for the stop-loss. This reduces losses from non-convergent trades.

Factor-neutral pair selection. We pre-filter pairs to ensure they are not simply co-loading on common Fama-French factors. Specifically, We require that the pair's spread has a correlation below 0.3 with each of the five Fama-French factors (MKT, SMB, HML, RMW, CMA). This dramatically reduces the number of qualifying pairs but ensures that the remaining pairs have a more fundamental economic basis for cointegration.

Table 2: Performance of adaptive methods over the 2020–2024 evaluation period.

The combined adaptive approach recovers a Sharpe ratio of 0.71 from the baseline of 0.24. The factor-neutral filter makes the largest single contribution by eliminating pairs whose cointegration is driven by common factor exposure rather than economic fundamentals. However, it also reduces the pair universe from 42 to 14, limiting portfolio diversification. The CUSUM structural break detector provides the second-largest improvement by cutting losses from non-convergent trades.

6. Is It Worth the Effort?

A Sharpe ratio of 0.71 from the adaptive approach is respectable but no longer exceptional. It compares unfavourably to the 1.2+ achieved by the classic approach in the 2000s and is roughly comparable to a simple time-series momentum strategy applied to the same universe. The question for practitioners is whether the additional complexity of cointegration testing, structural break detection, and factor-neutral filtering justifies the marginal performance.

Our view is that cointegration-based pairs trading in US equities is no longer a standalone strategy — the alpha has been largely arbitraged away. However, it remains useful as a component in a multi-strategy portfolio, particularly when combined with other market-neutral approaches. The adaptive methods We describe are necessary for the strategy to remain viable at all; without them, the classic approach is effectively random.

Opportunities may be better in less crowded markets. Commodity futures, emerging market equities, and cryptocurrency pairs show stronger and more persistent cointegration relationships, likely because these markets have fewer systematic arbitrageurs. We note that recent research on cryptocurrency pairs trading using the BTC-ETH pair has reported Sharpe ratios above 2.0, though these results should be discounted for the higher transaction costs and shorter track records available in crypto markets.

7. Conclusion

Cointegration-based pairs trading in US equities has experienced a secular decline in profitability, with the annualised Sharpe ratio falling from 1.24 to 0.19 over two decades. The decline is driven by increasing structural break frequency, strategy crowding, and spurious cointegration from common factor exposure. Adaptive methods — rolling re-estimation, real-time structural break detection, and factor-neutral pair selection — recover approximately two-thirds of the lost alpha, but the strategy is no longer the reliable profit source it once was. Practitioners should treat pairs trading as a diversifying component rather than a core strategy, and should look to less crowded markets for the strongest opportunities.