CMStatistics 2023: Start Registration
View Submission - CFE
A0870
Title: Bi-objective cost-sensitive machine learning: Predicting stock return direction using option prices Authors:  Robert James - The University of Sydney (Australia) [presenting]
Artem Prokhorov - University of Sydney (Australia)
Abstract: Cost-sensitive loss functions are studied for training machine learning models that predict the direction of future equity market index movement. In particular, a bi-objective loss function is designed that combines the log-loss with a second objective which asymmetrically penalizes individual false-positive and false-negative miss-classification errors. It is further discussed on how put and call option prices are natural measures of the misclassification costs. Using a comprehensive suite of classification performance metrics, it is investigated how training a linear elastic-net logistic regression model and a non-linear gradient boosting model using the cost-sensitive loss functions improves return direction predictions. A long/short investment strategy that uses the predictions from the cost-sensitive models improves risk-adjusted investment performance and reduces downside risk.