Uncertainty Awareness for Predicting Noisy Stock Price Movements

Abstract

Predicting stock price movements is challenging because financial markets are noisy – signals and patterns in different periods are dissimilar and often conflict with each other. Consequently, irrespective of whether the price rises or falls, none of the previous methods achieve high prediction accuracy in this binary classification task. In this study, we consider aleatoric uncertainty and model uncertainty when training neural networks to forecast stock price movements. Specifically, aleatoric uncertainty is known as statistical uncertainty. It indicates that similar historical price trajectories may not lead to similar future price movements. On the other hand, model uncertainty is caused by the model’s mathematical structures and parameter values, which can be used to estimate whether the models are familiar with the testing sample. Considering that most of the existing uncertainty estimation methods focus on model uncertainty, we transform the aleatoric uncertainty in financial markets to model uncertainty by removing samples with similar historical price trajectories and different future movements. The Bayesian neural network is then adopted to estimate the model uncertainty during inference. Experiment results demonstrated that the networks achieved high accuracy when they were certain about their predictions.