Economic Forecasting: Challenges and Neural Network Solutions

Macroeconomic forecasting is a very difficult task due to the lack of an accurate, convincing model of the economy. The most accurate models for economic forecasting, “black box” time series models, assume little about the structure of the economy. Constructing reliable time series models is challenging due to short data series, high noise levels, nonstationarities, and nonlinear effects. This paper describes these challenges and surveys some neural network solutions to them. Important issues include balancing the bias/variance tradeoff and the noise/nonstationarity tradeoff. The methods surveyed include hyperparameter selection (regularization parameter and training window length), input variable selection and pruning, network architecture selection and pruning, new smoothing regularizers, and committee forecasts. Empirical results are presented for forecasting the U.S. Index of Industrial Production. These demonstrate that, relative to conventional linear time series and regression methods, superior performance can be obtained using state-of-the-art neural network models. 1 Challenges of Macroeconomic Forecasting Of great interest to forecasters of the economy is predicting the “business cycle”, or the overall level of economic activity. The business cycle affects society as a whole by its fluctuations in economic quantities such as the unemployment rate (the misery index), corporate profits (which affect stock market prices), the demand for manufactured goods and new housing units, bankruptcy rates, investment in research and development, investment in capital equipment, savings rates, and so on. The business cycle also affects important socio-political factors such as the the general mood of the people and the outcomes of elections. The standard measures of economic activity used by economists to track the business cycle include the Gross Domestic Product (GDP) and the Index of Industrial Production (IP). GDP is a broader measure of economic activity than is IP. However, GDP is computed by the U.S. Department of Commerce on only a quarterly basis, while Industrial Production is more timely, as it is computed and published monthly. IP exhibits stronger cycles than GDP, and is therefore more interesting and challenging to forecast. (See figure 1.) In this paper, all empirical results presented are for forecasting the U.S. Index of Industrial Production. Macroeconomic modeling and forecasting is challenging for several reasons: No a priori Models: A convincing and accurate scientific model of business cycle dynamics is not yet available due to the complexities of the economic system, the impossibility of doing controlled experiments on the economy, and non-quantifiable factors such as mass psychology and sociology that influence economic activity. There are two main approaches that economists have used to model the macroeconomy, econometric models and linear time series models: Econometric Models: These models attempt to model the macroeconomy at a relatively fine scale and typically contain hundreds or thousands of equations and variables. The model structures are chosen by hand, but model parameters are estimated from the data. While econometric models are of some use in understanding the workings of the economy qualitatively, they are notoriously bad at making quantitative predictions. Linear Time Series Models: Given the poor forecasting performance of econometric models, many economists have resorted to analyzing and forecasting economic activity by using the empirical “black box” techniques of standard linear time series