Predictive Uncertainty in Neural Network-Based Financial Market Forecasting

In financial market forecasting, various methods based on statistical analysis and neural networks have been proposed. Accurate forecasting of future market states can be helpful in decision-making related to investment behavior; however, existing forecasting methods have considerable deficiencies due to the nature of financial markets and their complexity, influenceability, and nonstationarity. Forecasting of complex systems, such as financial markets, should be performed considering predictive uncertainty, and decision-making needs to be adjusted accordingly. In the present study, we introduce the concept of uncertainty to neural network-based financial market forecasting. A sparse variational dropout Bayesian neural network (SVDBNNs) is used for stochastic prediction, and on this basis, the corresponding decision-making process is proposed. The proposed method is validated by conducting investment simulation on the historical orderbook data from the Tokyo Stock Exchange and is confirmed to enable more efficient and safe investments compared with the considered alternative approaches.