Univariate Time Series Forecasting with Artificial Neural Networks: Relative Performance in a Large-Scale Macroeconomic Forecasting Experiment

Abstract

Artificial neural networks have proven to be highly effective tools in a variety of applications such as facial recognition, speech processing, medical diagnostics and more. In this paper, I test whether artificial neural networks can also be used to accurately forecast economic data in real-time. To do so, I conduct a forecasting experiment that compares the univariate forecasts of several linear and non-linear forecasting techniques at one, three, six and twelve month horizons for 100 monthly macroeconomic time series. Each forecast is produced in "real-time," using only data available up to the forecast date. Each forecasting method is compared against a benchmark no-change forecast and AR(6) forecast using three forecast comparison techniques, two of which are novel methods. Forecasts produced by linear forecasting methods tend to have lower out-of-sample mean-square errors than forecasts produced by artificial neural networks at each horizon. In multi-step ahead forecasts, other non-linear forecasting techniques such as K-nearest neighbor regression tend to outperform linear forecasting methods and artificial neural networks out-of-sample.