Empirical Analysis of Optimal Hidden Neurons in Neural Network Modeling for Stock Prediction

The neural network model presents a new procedure and structure in nonlinear interpolation and demonstrates a potential of stock market prediction with incomplete, imprecise and noisy data. However, a random selection of model parameters might cause the ¿over-fitting¿ or ¿under-fitting¿ problem in generalization or prediction. This paper presents a sensitivity analysis of optimal hidden layers and hidden neurons in neural network modeling for stock price prediction. To perform the sensitivity analysis, forty cases with various hidden neurons are examined to estimate the training and the generalization errors. The network architecture, the input pattern, and the complexity of problem are also taken into account in each case. The result with the minimum estimated generalization error is determined as the optimum for the application of neural network model.