Mixed Frequency Functional VARs for Nowcasting and Structural Analysis

We propose a functional-Vector Autoregressive model (fVAR) to nowcast the dynamics of the whole income distribution in the United Kingdom. British survey data about household income are published by the Office for National Statistics (ONS) with considerable delay, making them un- appealing for policy evaluation. Our approach produces accurate predictions of past, current and future income distributions. We introduce a framework to rank the predictive ability of forecasting models when the target object is a full density, rather than a single realization. Based on this novel loss-function, we establish that out fVAR provides superior forecasting accuracy with respect to com- peting models. Our model further allows to carry out structural analysis on the income distribution within a traditional VAR setting.