Benjamin W. Wah and Minglun Qian, University of Illinois, Urbana-Champaign
In this paper, we develop new constrained artificial-neural-network (ANN) formulations and learning algorithms to predict future stock prices, a difficult time-series prediction problem. Specifically, we characterize stock prices as a non-stationary noisy time series, identify its predictable low-frequency components, develop strategies to predict missing low-frequency information in the lag period of a filtered time series, model the prediction problem by a recurrent FIR ANN, formulate the training problem of the ANN as a constrained optimization problem, develop new constraints to incorporate the objectives in cross validation, solve the learning problem using algorithms based on the theory of Lagrange multipliers for nonlinear discrete constrained optimization, and illustrate our prediction results on three stock time series.There are two main contributions of this paper. First, we present a new approach to predict missing low-pass data in the lag period when low-pass filtering is applied on a time series. Such predictions allow learning to be carried out more accurately. Second, we propose new constraints on cross validation that can improve significantly the accuracy of learning in a constrained formulation. Our experimental results demonstrate good prediction accuracy in a 10-day horizon.