Neural network models with different input: An application on stock market forecasting

It is no doubt challenging to forecast the stock market accurately in reality due to the ever-changing market.  Ever since Artificial Neural Networks (ANNs) have been recognized as universal approximators, they are extensively used in forecasting albeit not having a systematic approach in identifying optimal input.  The appropriate number of significant lags of a time series corresponds to the optimal input in time series forecasting.  Hence, this study aims to compare the effect of several approaches in determining the input lag for ANNs prior to stock market forecasting, based on the autocorrelation function, the partial autocorrelation function, the Box–Jenkins model and forward selection.  The forecast performances of the ANNs were compared with benchmark models, namely the naïve and Box–Jenkins models, in terms of error magnitudes and trend change error.  In this study, all ANNs were found to outperform the benchmark models such that the neural network model trained with lags selected from forward selection of lag 1 and lag 31 (ANN4) is the best model as it achieved the highest accuracy with the lowest mean absolute percentage error and mean absolute error.  Contrary to expectations, all models performed poorly in forecasting the trend change of the stock series.  The ANNs with different inputs are viable in quantitative stock market forecasting yet more research is required to better understand other trend change measurements and improve the performance of forecasting the trend change of stock market.