Learning expectations using multi-period forecasts

This study investigates the macroeconomic implications of introducing perpetual learning in terms of multi-period forecasts to a simple search and matching model, to account for the model's lack of amplification and propagation of shocks. The model can match the amplification for vacancies and unemployment in the US data from 1955:Q1 to 2010:Q4 at the expense of deteriorating its predictions on autocorrelations and the slope of the Beveridge curve. The model with constant gain of 0.0045 can boost the amplification of the standard model by at least 50% while keeping correlations relatively unchanged. Adjustment costs in vacancies can improve the tradeoff between greater amplification and better correlations at a higher constant gain of 0.0095. At this gain the model can match the amplification in the data while maintaining the same correlations as the rational expectations model. Learning with decision rules that incorporate multiperiod forecasts, besides being consistent with the household's belief system, it produces autocorrelations for agents’ forecasting errors similar to those encountered in the survey of professional forecasters (1968:Q1–2015:Q2), while rational expectation and short horizon forecasting models imply a near zero autocorrelation for simulated forecasting errors.