A0441
Title: Bayesian dynamic graphical models for large vector autoregressions with time-varying parameters and volatility
Authors: Seyma Vahap - Kings College London (United Kingdom) [presenting]
Abstract: A proposed Bayesian dynamic graphical modelling (BDGM) approach has an important property of splitting a complex and high-dimensional vector autoregressive model with time-varying parameters and volatility discounting into locally structured sparse components. This high-dimensional model has an extensive collection of predictors, most of which make a small contribution to the overall power of the model. Still, it is unclear which predictors are relatively more important. A key role for local model specification and computations is the idea of pairwise conditional independence structure. This approach is achieved by developing an efficient Bayesian graphical variable selection method that can be applied recursively in parallel using a Gray code algorithm. The BDGM approach is applied to ten quarterly U.S. macroeconomic and financial variables. The results of posterior model probabilities over the space of all competing models show that there is considerable model uncertainty within the best-selected models. Then, a Bayesian model averaging (BMA) is performed to forecast the variables of interest over a pseudo out-of-sample forecast period 1984:Q2-2022:Q3. Comparing out-of-sample forecast performances shows that the joint model with BMA outperforms the joint model with the highest posterior probability over the majority of forecast horizons.
