A1224
Title: Market timing with bi-objective cost-sensitive machine learning
Authors: Artem Prokhorov - University of Sydney (Australia) [presenting]
Abstract: The aim is to develop a framework for cost-sensitive training of machine learning models that predict the direction of aggregate stock returns. A bi-objective loss function is designed that augments the traditional log-loss objective with an objective that minimizes the cost of individual false-positive and false-negative classification errors. It is argued that the option-implied conditional value-at-risk is a natural measure of the misclassification costs in such models. The bi-objective optimization framework permits us to isolate the effect of cost-sensitivity from log-loss minimization, and to integrate forward-looking information from options markets directly into the model training process. Changes are studied in the classification performance of elastic-net logistic regression and gradient-boosted decision trees trained using the bi-objective framework. The new approach improves the risk-adjusted returns of market timing strategies and substantially reduces downside risk.
