P artially Recurrent Neural Networks in Stock Forecasting
Dieter Merkl · 1996
Stock market data represent time-series data par excellence. The challenge for neural networks in such an environment is the representation of the temporal organization inherent in the input data. Conventionally , the temporal organization is approximated with a sliding windows technique when using standard feedforward neural networks such as the Multi-Lay er Perceptron as the underly ing model during the learning process. This approach suffers from a rather random selection of the sliding windows's actual size. We show that a more compact representation can be achieved by using partially recurrent neural networks. Moreover, the forecasting results as the ultimate goal of the application are improved significantly. 1. INTRODUCTION The application of artificial neural networks has a long tradition in stock forecasting. The reasons are quite obvious when taking into consideration the features of artificial neural networks such as being robust against noisy input data or their ability of...