Title: Density forecasts of inflation and interest rates using Bayesian DCS models with dynamic conditional skewness
Authors: Blazej Mazur - Cracow University of Economics (Poland) [presenting]
Abstract: The focus is on potential gains in density forecasting performance of univariate macroeconomic series using models that allow for dynamic changes in conditional asymmetry. The conditional distribution used is flexible as it allows for skewness and shape (tail) asymmetry, being a two-piece asymmetric generalized $t$. The benchmark model assumes the Dynamic Conditional Score mechanism for updating of conditional location and scale, with asymmetry/skewness parameters being time-invariant. It is then extended using two alternative parametrizations implying dynamic conditional skewness. Within the first approach the overall scale parameter is separated from the skewness coefficient, whereas the other one is based on left-scale and right-scale parameters (with overall scale and skewness coefficients being implicit). Within the two approaches diagonal and matrix versions of DCS models are compared. The matrix specification allows for dependencies in dynamics of the three features (location, scale and skewness or location, left-scale and right-scale). The aim is to identify optimal parametrization and to indicate crucial dynamic dependencies. Within the forecasting experiment, disaggregate US PCE inflation rates and interest rates are analyzed. Density forecasting performance is examined using CRPS and log-Score criteria, calibration is investigated using PITs; performance of linear pools is also considered.