Modelling and Forecasting Unemployment Non-linear Dynamics Using Spectral Analysis

Changes in the unemployment in Croatia are largely permanent. However, transitory movements account for most of the unemployment dynamics after 2008. Unemployment series is non-stationary but mean-reversing, non-linear with structural breaks and significant white and red-noise in the data. This paper estimates Multivariate singular spectrum model (MSSA) to explain overall fluctuations in unemployment registered during 1998-2013. Unemployment behaviour in Croatia shows evidence of cyclo-stationarity caused by seasonal employment effects. We use 88 time series (variable) to explain observed fluctuations with our MSSA model explaining 76 % of the total unemployment variance comprehensively. Evidence of this study demonstrates that unemployment phenomena should be modelled by using a non-linear model with multivariate singular spectrum models giving more robust and empirically valid results in relation to standard modelling techniques. A 5-6 years limit cycle for unemployment is isolated dominating unemployment behaviour in Croatia over the last two decades.