CONSTRAINED FORMULATIONS AND ALGORITHMS FOR PREDICTING STOCK PRICES BY RECURRENT FIR NEURAL NETWORKS

In this paper, we develop a new constrained artificial-neural-network (ANN) formulation and the associated learning algorithm for predicting stock prices, a difficult time-series prediction problem. We characterize daily stock prices as a noisy non-stationary time series and identify its predictable low-frequency components. Using a recurrent finite-impulse-response ANN, we formulate the learning problem as a constrained optimization problem, develop constraints for incorporating cross validations, and solve the learning problem using algorithms based on the theory of extended saddle points for nonlinear constrained optimization. Finally, we illustrate our prediction results on ten stock-price time series. Our main contributions in this paper are the channel-specific low-pass filtering of noisy time series obtained by wavelet decomposition, the transformation of the low-pass signals to improve their stationarity, and the incorporation of constraints on cross validation that can improve the accuracy of predictions. Our experimental results demonstrate good prediction accuracy and annual returns.