Pairs trading with wavelet transform

Abstract

We show that applying the wavelet transform to S&P 500 constituents' prices generates a substantial increase in the returns of the pairs-trading strategy. Pairs trading strategy is based on finding prices that move together, but if there is shared noise in the asset prices, the co-movement, on which one base the trades, might be caused by this common noise. We show that wavelet transform filters away the noise, leading to more profitable trades. The most notable change occurs in the parameter estimation stage, which forms the weights of the assets in the pairs portfolio. Without filtering, the parameters estimated in the training period lose relevance in the trading period. However, when prices are filtered from common noise, the parameters maintain relevance much longer and result in more profitable trades. Particularly, we show that more precise parameter estimation is reflected on a more stationary and conservative spread, meaning more mean reversion in opened pairs trades. We also show that wavelet filtering the prices reduces the downside risk of the trades considerably.

Keywords:

• Pairs trading
• Wavelet transform
• Minimum distance method
• Cointegration method
• Statistical arbitrage

JEL Classification:

• C1
• C22
• C53
• G11
• G12
• G14

Disclosure statement

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

Supplemental data

Supplemental data for this article can be accessed online at http://dx.doi.org/10.1080/14697688.2023.2230249.

Notes

1 We find a negative correlation between the noise we separate and the returns to standard methods of unfiltered pairing.

2 We refer the readers to the online appendix mentioned in the introductory section of this study.

3 Our dataset does not cover any large crisis periods, yet our claim is that the removal of common market noise prior to pairing will particularly be relevant for turbulent times where all prices tend to move together and lucrative pairs are difficult to identify.

4 The solution of what happens to parameter variances is straightforward if the prices were to be linearly related; however our moving window (year-by-year) estimation hints towards contrary, suggesting a non-linear relation between two price series.

5 The detailed description of the simulation is provided in the online appendix.

6 We note that the case ${ϵ}_{t_{o,m,n},w}<0$ leads to the same specification, since trades involved are the same. Therefore, we only report the specification when ${ϵ}_{t_{0,m},w}>0.$,