43 Forecasting Aggregate Sales with Interest Rates Using Multiple Neural Network Architectures

This paper examines out-of-sample forecasts of aggregate sales using 3-month treasury bills interest rate in NeuroSolutions environment referenced against forecasts of linear regression models. Two types of dynamic neural network models trained with the Levenberg-Marquardt backpropagation algorithm under supervised learning were used to investigate the hypothesis that aggregate manufacturing and trade sales are positively related to inverted interest rates and that inverted interest rates are an efficient predictor of the future values of aggregate sales. Interest rates and aggregate sales are important macroeconomic variables used as business cycle indicators. Interest rates are a lagging economic indicator while inverted interest rates are a leading one. Manufacturing and trade sales are a coincident indicator. Seven performance metrics were used to measure the relative efficacy of the predictions with the neural network models out-performing the linear regression models and the Elman models producing smoother forecasts and better performance on the sectionalized data.