Machine Learning, Classification Algorithm and Cross Section of Stock/Bond Returns

This paper compares models trained with classification algorithm with those trained using regression algorithm. We use technical indicators, 43 bond characteristics and 94 stock characteristics as predictors. We find very strong evidences of the superior predictive ability of classification models over regression models for bonds. All machine learning models for classification algorithm have statistically significant mean returns, unlike regression algorithm, in which only 3 out of 10 models are statistically significant. classification models have significantly better results for the evaluation metric; mean bond return, alpha intercept, cumulative return, Sharpe ratio, maximum drawdown, zero return BETC and insignificant BETC. For stocks, we do not find strong evidences of the predictive advantage of classification models over regression model. For bonds, classification models are able to exploit the information contained in bond characteristics besides technical indicators, unlike regression models, therefore leading to significantly better results