When Knowing Less Fits More: Comparing Rational Expectations and Adaptive Learning in DSGE Models

Sammanfattning

A central feature of modern macroeconomics is that agents form expectations about future economic variables under rational expectations (RE). However, evidence suggests agents form expectations based on more limited information than RE assumes. This thesis therefore compares the empirical performance of a DSGE model under RE against one with adaptive learning (AL). Under AL, agents form expectations using a univariate AR(2) forecasting model and update their beliefs recursively using a Kalman Filter. We estimate both the RE and AL versions of the model using Bayesian methods on quarterly euro area data from the Area-Wide Model Database covering 1970-2019. Our results indicate that the AL specification fits the euro area data substantially better than the RE benchmark. Adaptive learning generates endogenous persistence by making agents more backward-looking, reducing the model's reliance on mechanical price stickiness and yielding more realistic propagation of monetary policy shocks. Interestingly, the estimated degree of active belief updating is small, suggesting that the empirical gains stem primarily from the bounded rationality embedded in rule-of-thumb AR(2) forecasting structure rather than from the Kalman Filter learning itself.