Improving Pairs Trading with Cluster-Based Pair Selection

Pairs trading is a classic quantitative trading strategy. Despite its widespread use, it continues to attract research attention. A recent line of research focuses on grouping underlyings into clusters with similar characteristics. We recently discussed such grouping using fundamental metrics.

Reference [1] also attempts to improve the pair selection process, but this time using a purely quantitative, unsupervised machine learning approach. Specifically, it groups the underlyings using density-based methods applied to the correlation-distance matrix. The objective is to identify neighborhoods of underlying that exhibit similar residual return dynamics, that is, comparable co-movement patterns after removing broad market effects, so that cointegration tests and mean-reversion diagnostics are applied primarily within economically coherent groups.

The authors pointed out,

The performance we obtain is encouraging given the constraints of our design: we use only daily adjusted closes, transparent mean-reversion filters, and simple z-score entry/exit rules without any intraday microstructure signals. Across the four clustering specifications in Table 2, annualized Sharpe ratios range from 1.7228 to 2.4935. The best risk-adjusted variant is OPTICS with PC1 removed (Sharpe 2.4935), which also achieves the smallest maximum drawdown (-8.74%). At the same time, the highest terminal wealth is achieved by HDBSCAN with PC1 removed (cumulative return 1020.34%), albeit with materially larger downside risk (max drawdown -28.39%). Overall, the results suggest that a fully reproducible, daily-data pipeline can still recover substantial relative-value structure in a broad ETF universe.

A natural benchmark is a passive investor who buys and holds a broad equity index over the same window. Over 2014–2025, Table 2 shows that SPY (S&P 500) delivers a cumulative return of 302.03% (wealth multiplier 1 + 3.0203 = 4.0203) while QQQ (Nasdaq–100) delivers 542.53% (wealth multiplier 1 + 5.4253 = 6.4253). By contrast, our pairs trading specifications produce cumulative returns between 449.48% and 1020.34%, corresponding to wealth multipliers between 5.4948 and 11.2034…

In summary, this paper shows that a simple, transparent pairs trading framework applied to ETFs using daily data and clustering techniques can achieve strong risk-adjusted performance, with Sharpe ratios between 1.72 and 2.49. The strategy produces cumulative returns between 449% and 1020%, outperforming passive benchmarks such as SPY and QQQ over the 2014–2025 period.

We find the results noteworthy, particularly given that the method is applied to liquid ETFs. As is well known, compared with single-name stocks, ETFs typically carry lower risk and therefore lower expected returns. Nevertheless, the proposed method delivers strong risk-adjusted returns without the use of leverage.

Let us know what you think in the comments below or in the discussion forum.

References

[1] Fuwen Gan and Ramy Mizrachi, Cluster-Based Pairs Trading: Combining Unsupervised Learning with Cointegration Filtering, 2025, Rice University

Post Source Here: Improving Pairs Trading with Cluster-Based Pair Selection

source https://harbourfronts.com/improving-pairs-trading-cluster-based-pair-selection/