Measuring economic cycles in data

Abstract

To determine if enough data is available for forecasting or stress testing, a better measure of data length is required. The number of months or years of history we collect does not capture the autocorrelation structure of the data, and higher frequency sampling does not increase information content in most cases. Measures such as effective sample size provide ways of measuring data length but with recognized flaws. Analysts often resort to citing the number of recessions in the data, but this does not give a true measure of the amount of structure in data series from countries such as Australia or China. To overcome these limitations, the current work proposes a data driven metric that measures the number of economic cycles weighted by the severity of those cycles, utilizing a state space reconstruction approach. This measure can be employed to describe the amount of useful structure in historic data from a modeling perspective; current alternatives provide no such metric. Having an objective measure of economic cycles can be an important element of model risk management. Data length is a potential risk and can be a component of model risk scoring for corporate model inventories.