Title: Identification of pairs trading opportunities using copula-based conditional probabilities and machine learning models
Authors: Ting Hin Cheng - The Open University of Hong Kong (Hong Kong) [presenting]
Ka Lok Li - The Open University of Hong Kong (Hong Kong)
Chun Lam Wong - The Open University of Hong Kong (Hong Kong)
Suet Ying Lau - The Open University of Hong Kong (Hong Kong)
Chak Long Ng - The Open University of Hong Kong (Hong Kong)
Carlin Chu - The Open University of Hong Kong (Hong Kong)
Abstract: Copula is a multivariate probability distribution function that allows a separate estimation of marginal and joint distributions for modeling relationships between variables. It is flexible to capture non-symmetric relationships between the tails and this makes it a superior candidate for financial applications. In this study, the conditional probabilities of copula are modeled as input features for the identification of pairs trading investment opportunities. Apart from the usage of traditional logistic regression model, 3 machine learning models (Neural Network, AdaBoost and Random Forest) are employed to explore the corresponding contributions under various settings. High frequency (1-minute interval) trading prices during the period of October 2018 to February 2019 are collected from Bloomberg terminal for carrying out the empirical analysis of this study. To come up with a more comprehensive picture, several distinct settings (i.e. under-sampling, cross-validation, early stopping, dropout and grid search of hyper parameters) are investigated in this study. Our empirical results show that logistic regression of copula-based features does not work well. However, the generated features help to produce favorable performances when they are used together with sophisticated machine learning models. The AdaBoost model produces the highest ROC index while the 5-layer Neural Network model delivers the largest profit.