|
Abstract:
|
It is demonstrated that hedging portfolio gains can be achieved using Machine learning (ML) methods in finance. However, the performance of many ML methods is sensitive to choices of hyperparameter values. In this paper, we focus on the ML methods for regression and classification in a panel of stock returns data with hyperparameters tuned automatically through Gaussian process (GP), which is a popular surrogate model for Bayesian optimization, a derivative-free approach for optimization of black-box functions. In simulation studies, we validate accuracy of our methods in predicting stock return means and classifications. We also find hedging portfolios constructed based on predicted classification probabilities from ML methods are more profitable than traditional regression portfolios when data are noisy and highly skewed. An application in the US stock market is also provided.
|
