In search of pairs using firm fundamentals: is pairs trading profitable?

Abstract

We investigate whether spurious pairs created by multiple hypothesis testing can be minimized when firm characteristics are used to identify pairs. The results show that the portfolios of pairs that have higher similarities in firm characteristics outperform those that are fundamentally less similar by minimizing the non-convergence risk. The cost of trading spurious pairs is significant despite the empirical results that the profitability of pairs trading has continued to decline since 2003 and is not significant anymore. Accounting information plays a crucial role in identifying pairs rather than market trading data, and the importance of firm fundamentals in pairs trading increases during market crises.

KEYWORDS:

• Elastic net regression
• multiple hypothesis testing problem
• fundamental similarity
• disappearance of pairs trading profitability

JEL classifications:

• G11
• G12
• G14

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 We have tested artificial neural networks and other methods using machine learning. However, the elastic net regression appears to be the most useful for summarizing firm fundamentals.

2 Studies regarding machine learning have recently been conducted in various areas of finance. For example, Moritz and Zimmermann (2016), Freyberger, Neuhierl, and Weber (2020), and Gu, Kelly, and Xiu (2020) demonstrated that machine-learning techniques outperform conventional methodologies in predicting asset prices. These studies suggest that non-linearity and big data are better handled by machine-learning techniques than conventional model-based asset pricing methodologies.

3 Firm characteristics with missing values of more than 40% during a specific period are excluded from the sample for that period. Consequently, firm characteristics are time varying. Additionally, if either ${\text{FC}}_{k,t-1}^{p1}{\text{or FC}}_{k,t-1}^{p2}$ has missing values during preprocessing, ${\text{ADFC}}_k^i$ is considered to be a missing value, which is subsequently replaced by the mean.

4 The regression is estimated with approximately 6000 samples (60 months × 100 pairs).

5 See the Appendix for additional details.

6 The OLS method estimates ${\widehat{\text{β}}}_{k,t}$ by using ordinary least squares (OLS) instead of elastic net in Equation (4) and then follows the empirical analysis methodology outlined in Section 2.

7 The results can be obtained from the authors upon request.

8 Jacobs and Weber (2015) demonstrate that pairs trading in emerging markets, which be known to be more inefficient than developed markets, yields returns that are about a quarter higher than in developed markets.

Additional information

Notes on contributors

Sungju Hong

Sungju Hong received his master degree in economics (2022) from Sungkyunkwan University (SKKU). Currently, he works as an index researcher and system manager for portfolio performance analysis in FnGuide, Seoul. His research interest is in the area of finance and machine learning.

Soosung Hwang

Soosung Hwang received his PhD in the Faculty of Economics (1997) from the University of Cambridge. He is professor of finance in the department of economics, Sungkyunkwan University (SKKU). His teaching activities cover asset pricing, financial econometrics, and machine learning. His research topics are in the areas of empirical asset pricing, financial econometrics, behavioral finance, real estate prices, and machine learning.