Forecasting and Signal Extraction with Misspecified Models

The paper illustrates and compares estimation methods alternative to maximum likelihood, among which multistep estimation and leave-one-out cross-validation, for the purposes of signal extraction, and in particular the separation of the trend from the cycle in economic time series, and long-range forecasting, in the presence of a misspecified, but simply parameterised model. Our workhorse models are two popular unobserved components models, namely the local level and the local linear model. The paper introduces a metric for assessing the accuracy of the unobserved components estimates and concludes that cross- validation is not a suitable estimation criterion for the purpose considered, whereas multistep estimation can be valuable. Finally, we propose a local likelihood estimator in the frequency domain that provides a simple and alternative way of making operative the notion of emphasising the long-run properties of a time series.