Analysis of pairwise cointegration in the US stock market

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
School of Business | Master's thesis
Degree programme
In this thesis I introduce and evaluate the empirical performance of a Bayesian cointegration testing procedure. In contrast with previous Bayesian cointegration tests, the introduced procedure takes into account cross sectional variation in the prior probability of cointegration, and models this variation as a function of similarity metrics between company attributes. The procedure is motivated both by arbitrage pricing theory, as well as previous empirical findings on the attributes of companies with similarly behaving stock prices. I analyze the daily closing prices of a sample of US-listed common stock during a ten-year period between 2009 and 2018. The former half of this sample is used for analyzing the cross sectional variation in the probability of pairwise cointegration for non-overlapping 12 month periods. This variation is explained by a set of similarity metrics based on the industry, book-to-market-ratios and market capitalizations of companies. I show that similarities between two companies’ industries and book-to-market-ratios positively affect the probability of pairwise cointegration. In contrast, similarity with respect to market capitalization is negatively associated with the probability of cointegration. I derive a logistic model for computing conditional prior odds of cointegration based on the similarity of company attributes, where the conditionality is with respect to an unobservable parameter representing the average cointegration probability. Using the latter half of the sample, I test the ability of the introduced procedure in identifying persistent pairwise cointegrating relations. I classify pairs as cointegrated based on conditional posterior odds after the first six months of every 12-month period, and subsequently test whether the pair persists to be classified as cointegrated during the remaining six months. I show that the procedure is able to identify persistent pairwise cointegration even before the introduction of conditional prior odds. Further, when conditional prior odds are introduced, the empirical performance of the procedure is improved, especially when the classification threshold is high. The research contribution of this thesis is threefold. Firstly, a conceptual contribution is made by introducing the concept of conditional prior odds, the first of its kind in pairwise cointegration testing between stocks. Secondly, the thesis offers an empirical contribution by analyzing the effects of company attributes on cointegration probability. Thirdly, the thesis offers practitioners an empirically tested procedure for the identification of persistent pairwise cointegration between stocks, that can be utilized in a pairs trading context.
Thesis advisor
Puttonen, Vesa
Bayesian, cointegration, pairs trading, statistical arbitrage, USA
Other note