The Treasury Bill Rate, the Great Recession, and Neural Networks Estimates of Real Business Sales

This paper analyzes out-of-sample forecasts of real total business sales. We study monthly data from January 1970 to June 2012. The predictor variable, 3-month Treasury bill interest rate, was used with both the regression (used as a benchmark) and neural network models. The neural network models’, trained in supervised learning with the Levenberg-Marquardt backpropagation through time algorithm, prediction accuracy was confirmed with correlation coefficient and root mean square tests. The activation function used for the focused gamma models of the time-lag recurrent networks in both the hidden and output layers was tanh. The forecast period ranged from January 2006 to June 2012 thus encompassing the past recession. The real business sales variable is one of the indicators used as a coincident index of the U.S. business cycle, and is included among the variables studied by the Federal Reserve to formulate monetary policy. It is thus an important indicator surrogating for real GDP, which is reported quarterly and with a longer time delay. Our analysis shows that recent recessions have increased in duration, so that using a 36-month change to approximate an average cycle in estimating and forecasting is more relevant and accurate than past usage of a 24-month change.