Finance & Financial Econometrics: 
Time: 11.30 am
Date: 11th of May 2023
Room 3001

Pierluigi VALLARINO (Aarhus University) “Time-Varying Kernel Densities as Dynamic Infinite Mixture Models”

Abstract : Building on kernel density estimation for time series data, we introduce the family of Dynamic Infinite Mixture Models (DIMMs). DIMMs approximate the time-varying distribution of a time series with that of an infinite mixture of location-scale random variables. Different specifications of a DIMM can capture different features of the time series of interest, such as different memory properties of the predictive mean and asymmetric effects in the predictive variance. A maximum likelihood estimator is proposed. Its asymptotic properties are studied under a fully misspecified setting and its finite sample behaviour is assessed in a Monte Carlo analysis. An application to US GDP growth shows that DIMMs: i) improve upon extant kernel density approaches for time series data; ii) reliably track the time-varying distribution of interest; iii) perform on par with – if not better than – a fully-fledged parametric model when it comes to predicting probability density functions.

Organizers:
Jean-Michel ZAKOIAN (CREST)

Sponsors:
CREST