On the Monitoring and Prediction of US Business Cycles using Statistical Process Monitoring

Journal Title
Journal ISSN
Volume Title

The Business Cycle (BC) paradigm of Mitchell and Burns has evolved from their original goal of understanding economic performance across the entire Business Cycle of expansions, recessions, contractions, and revivals to the binary identification of Business Cycle Turning Points (BCTPs), peaks and troughs. The literature related to the prediction of BCTPs includes univariate and multivariate methods ranging from heuristic rules for identifying TPs to advanced Markov Switching Models and machine learning algorithms. Across all the methods we see a fundamental issue related to the lack of understanding of economic performance that stems from i) structuring the BCTP prediction problem as a bivariate model and ii) from sampling variation not being accounted for when making decisions on changes in economic performance. We propose a new paradigm for modeling the Business Cycle which expands the binary economic state to a continuous characterization of aggregate economic activity and develops methods that incorporates sampling variation. The new paradigm is applied to the univariate problem of modeling Gross Domestic Product (GDP) to predict BCTPs and to the multivariate problem of predicting BCTPs using the four leading indicators the National Bureau of Economic Research (NBER) uses to heuristically predict BCTPs. This research aims to show that the new paradigm of a continuous model for economic performance and accounting for sampling variation can effectively predict BCTPs with an increase in the amount of information related to decisions regarding economic performance. The univariate approach is based on the Statistical Process Monitoring technique of Self-Starting Cumulative Sum (SSCUSUM) control charts. The SSCUSUM charts allow for the identification of patterns of changes in the mean and/or std. dev. of economic indicators across the entire economic process. A case study is conducted using real GDP % growth between 1965 and 2019, which shows the SSCUSUM control charts can: identify periods of consistent endogenous variation or steady state performance with statistically differentiable means and/or standard deviations, reliably reproduce the NBER BCTPs, identify patterns of economic activity leading up to and away from recessions, and identify over twice the amount of information on economic performance as the current bivariate methods. The SSCUSUM method identified 38 changes in the mean or standard deviation of real GDP % growth, while the NBER TPs identified 7 peaks and 7 troughs in economic performance over the same period. Multivariate methods of predicting BCTPs rely on the same bivariate model of the BC. The only definition this model has for economic performance comes from the BCTPs defined by the NBER. Literature related to the identification of BCTPs have typically used the average Time to Signal as a metric and analyzed a single, historically observed, multi-variate data paths. For the multivariate problem we will address the issue of selection of design parameters for detection algorithms. Using a Monte Carlo simulation, we show that using the new paradigm of a continuous response for economic performance and accounting for sampling variation allows a researcher to understand the ability of an algorithm to detect recessions of differing severities.

Self-Starting Cumulative Sum Control Chart, Machine Learning, Statistical Process Monitoring, Business Cycle, Learning Vector Quantization, Economic Performance