Optimizing NARX-RNN Performance to Predict Precious Metal Futures market

Abstract

Precious metals offer lucrative investments appealing to investors globally, leading to a surge in demand for accurate forecasts. Published literature for prediction applications often employs Artificial Neural Networks (ANNs), possessing desirable generalization over nonlinear data and design flexibility. Recurrent Neural Networks (RNNs) are a class of ANNs designed for time series forecasts providing superior approximations. Nonlinear Autoregressive with Exogenous input (NARX) is an RNN model with high memory retention properties, applied in this study to predict ten assets from the precious metal futures market, for three-month predictions (April 2021-June 2021). Network inputs are evaluated through feature selection to filter uncorrelated factors from the network dataset. Accuracy of prediction is enhanced through multi-objective Response Surface Methodology (RSM) optimization, as several variables characterize RNN performance. Three key variables are selected for analysis through RSM, providing optimum configuration to obtain targeted outcome. Simulation results reveal that five assets produce acceptable result, showing an improved fitness through RSM-suggested configurations. Observations indicate intercorrelation between RSM inputs, highlighting its efficiency over conventional methods. Implementing additional RSM inputs to develop more complex models might achieve further reliability. This research provides performance improvement measures for RNNs utilized in financial data projections.