Is unemployment asymmetric over the business cycle?

Abstract

In this study we revisit a significant work conducted by Rothman in 1998. With a view to finding the most suitable model for the interpretation of the asymmetric unemployment rates, various nonlinear ARMA models were introduced estimated and examined for their ability to account for this asymmetric behaviour plus forecast its future values. We aim on replicating the estimations, with data of the U.S. unemployment rates corresponding to the same as well as an extended time period, for the purpose of finding out how the models vary when we add additional and more modern observations. Data of the Greek unemployment rates is also used, demonstrating how these models look for this particular case and if there is a difference with the first one. Graphs showing the course of the coefficients values and t-statistics over time are also included. The models used are: Autoregressive (AR), Self-Exciting autoregressive (SETAR), Exponential smooth transition Autoregressive (ESTAR), Exponential Autoregressive (EAR), Generalized Autoregressive (GAR) and the Bilinear. We begin with the replication of the models using the same dataset, of the US quarterly unemployment rates of the period 1949:1 - 1979:4. The results indicate small deviations mainly on the thresholds of the two Threshold autoregressive models (TAR). Afterwards we expand the dataset to 2008:2 and repeat the estimations for the same models, this time finding only two models to be significant. Out of sample forecasts are then performed and interesting results are found for the HP-Filtered data. Lastly, we repeat the same analysis using Greek unemployment rates for the time period of 2001:1 - 2021:4 and conclude with similar results for the two countries, with the ESTAR model proving to be the best expansion of the Autoregressive model for both cases, even though it still needs a lot of improvement.