Learning Uncertainty in Market Trend Forecast Using Bayesian Neural Networks

Forecasting financial market trends is challenging. Predicting financial market trends always involves uncertainty because the economy is a complex system with a wide variety of interactions. Thus, to consider uncertainty, trends must be estimated stochastically. Conventional machine learning and deep learning methods cannot learn prediction uncertainty, and predicted “probabilities” calculated using such methods are unreliable. Bayesian neural networks (BNN) have been studied relative to their ability to consider uncertainty. A BNN enables Bayesian inference by considering a prior distribution of parameters in a neural network (NN). We have proposed a Bayesian convolutional neural network (CNN) method to predict short-term stock price trends using stock order data. We used a CNN to learn the series features of the stock order data and a dropout technique is employed to enable Bayesian inference. The proposed Bayesian CNN model trained prediction uncertainty properly, and the predicted probability calculated by the proposed model is more likely to predict the actual market movement than the conventional CNN and logistic regression models. It is expected that using Bayesian CNN will provide more reliable market trend forecasts.