The study reports empirical evidence that artificial neural network based models are applicable to forecasting of stock market returns. The Nigerian stock market logarithmic returns time series was tested for the presence of memory using the Hurst coefficient before the models were trained. The test showed that the logarithmic returns process is not a random walk and that the Nigerian stock market is not efficient. Two artificial neural network based models were developed in the study. These networks are TECH (4-3-1) and TECH (3-3-1) whose out-of-sample forecast performance was compared with a baseline ARIMA (3,0,1) model. The results obtained in the study showed that artificial neural network based models are capable of mimicking closely the log-returns as compared to the based model. The out-of-sample evaluations of the trained models were based on the RMSE, MAE, NMSE and the directional change metric respectively. Based on these metrics, it was found that the artificial neural network based models outperformed the ARIMA based model in forecasting future developments of the returns process. Another result of the study shows that instead of using extensive market data, simple technical indicators can be used as predictors for forecasting future values of the stock market returns given that the returns has memory of its past.
Artificial neural networks Long memory Random walk Forecasting Training Stock Market Returns Technical analysis indicator ARIMA.