Empirical analysis of parallel-NARX recurrent network for long-term chaotic financial forecasting

Financial data are characterized by non-linearity, noise, volatility and are chaotic in nature thus making the process of forecasting cumbersome. The main aim of forecasters is to develop an approach that focuses on increasing profit by being able to forecast future stock prices based on current stock data. This paper presents an empirical long term chaotic financial forecasting approach using Parallel non-linear auto-regressive with exogenous input (P-NARX) network trained with Bayesian regulation algorithm. The experimental results based on mean absolute percentage error (MAPE) and other forecasting error metrics shows that P-NARX network trained with Bayesian regulation slightly outperforms Levenberg-marquardt, Resilient back-propagation and one-step-secant training algorithm in forecasting daily Kuala Lumpur Composite Indices.