Explaining Unemployment Rates with Symbolic Regression

Abstract

Much of the research on the accuracy of symbolic regression (SR) has focused on artificially constructed search problems where there is zero noise in the data. Such problems admit of exact solutions but cannot tell us how accurate the search process is in a noisy real world domain. To explore this question symbolic regression is applied here to an area of research which has been well-travelled by regression modelers: the prediction of unemployment rates. A respected dataset was selected, the CEP-OECD Labor Market Institutions Database, to provide a testing environment for a variety of searches. Metrics of success for this paper went beyond the normal yardsticks of statistical significance to demand “plausibility”. Here it is assumed that a plausible model must be able to predict unemployment rates out of the sample period for six future years: this metric is referred to as the “out of sample R²”. We conclude that the two packages tested, Eureqa and ARC, can produce models that go beyond the power of traditional stepwise regression. ARC, in particular, is able to replicate the format of published economic research because ARC contains a high level Regression Query Language (RQL). This research produced a number of models that are consistent with published economic research, have in sample R² values over 0.80, no negative unemployment rates, and out of sample R² values above 0.45. It is argued that SR offers significant new advantages to social science researchers.