Direction-of-change financial time series forecasting using Bayesian learning for MLPs

Conventional neural network training methods find a single set of values for network weights by minimizing an error function using some gradient descent-based technique. In contrast, the Bayesian approach infers the posterior distribution of weights, and makes predictions by averaging the predictions over a sample of networks, weighted by the posterior probability of the network given the data. The integrative nature of the Bayesian approach allows it to avoid many of the difficulties inherent in conventional approaches. This paper reports on the application of Bayesian MLP techniques to the problem of predicting the direction in the movement of the daily close value of the Australian All Ordinaries financial index. Predictions made over a 13 year out-of-sample period were tested against the null hypothesis that the mean accuracy of the model is no greater than the mean accuracy of a coin-flip procedure biased to take into account non-stationarity in the data. Results show that the null hypothesis can be rejected at the 0.005 level, and that the West p-values obtained using the Bayesian approach are smaller than those obtained using conventional MLPs methods.