Links
- Paper (preprint)
Abstract
I evaluate the properties and performance of band-spectral estimators applied to business cycle models. Band-spectral methods are widely used to study frequency-dependent relationships between time series. In business cycle research, this approach permits the estimation of structural models on the basis only of the frequencies they are best suited to represent, such as the business cycle frequencies. In particular, the frequency domain approximation of the likelihood function (the Whittle likelihood) can be used to estimate the parameters of models on the basis of a targeted band of frequencies. Using the medium-scale model of Angeletos et al. (2018) as a data-generating process, I conduct a Monte Carlo study to evaluate the finite-sample properties of the band-spectral maximum likelihood estimator (MLE) and to compare them to those of the full spectrum and the exact time-domain MLE. The results show that using the band-spectral estimator leads to considerable biases and efficiency losses for most estimated parameters. In fact, the performance of both Whittle likelihood-based estimators is found to be seriously deficient in terms of bias and accuracy, in contrast to that of the time domain estimator, which successfully recovers all model parameters. I show how these findings can be explained with the theoretical properties of the underlying model, and describe simple-to-use tools and diagnostics that can be used to detect potential problems in band-spectral estimation for a wide class of macroeconomic models.