Essays in Empirical Macroeconomics: Identification in Vector Autoregressive Models and Robust Inference in Early Warning Systems

Abstract

Bayesian empirical macroeconomic models are excellent tools for prediction and structural analysis. The use of a prior distribution facilitates model averaging, allows for structural identification of multiple time series models and makes estimation of high-dimensional models feasible. However, prior distributions need to be chosen carefully in order to accurately reflect the researcher's beliefs before looking at the data. I exemplify how to do so in this thesis by employing model averaging techniques in Bayesian spirit, by developing tools to express priors for structural vector-autoregressive models, and by showing a new approach to identify the impact of variations in uncertainty in a data-intensive environment.

In the first chapter, which is based on joint work with Tigran Poghosyan, we use an Early Warning System (EWS) to recover leading indicators of fiscal distress events. In particular, we use Extreme Bounds Analysis (EBA), a model averaging approach introduced by Leamer (1985) and popularised by Sala-i-Martin (1997), to assess the robustness of leading indicators for fiscal distress across different models. We find that both fiscal and non-fiscal leading indicators are robust predictors of fiscal distress events. In a second step we assess the forecasting performance of an EWS based on the most robust leading indicators. We find that it offers a gain in predictive power compared to a baseline model which is based on fiscal leading indicators only. Lastly, we assess the robustness of these results across various model specifications, subsamples and estimation strategies.

In the second chapter, which is based on joint work with Michele Piffer, we propose a new approach to express prior beliefs on the impulse responses of structural vector auto-regressive (SVAR) models. This approach does not restrict the family of prior distributions to a set that is technically convenient. Rather, it combines extensive flexibility in the choice of priors with an efficient importance sampler to explore the posterior distribution. We compare our new posterior sampler to a computationally more demanding generic sampler and confirm that we recover the shape of the posterior. We illustrate the approach using artificial data and in an application of sign restrictions to identify oil market shocks. We show that posterior inference is sharpened compared to the traditional approach of imposing sign restrictions and that oil supply shocks play a major role in driving oil price dynamics.

In the third chapter I investigate the effects of uncertainty shocks in the spirit of Bloom (2009) using a newly developed Bayesian Proxy Factor-augmented VAR (BP-FAVAR) model. This model combines a large information set with an identification scheme based on external instruments, thereby jointly addressing informational deficiency issues and non-credible identification assumptions. I propose a new sampling algorithm exploiting the state-space representation of the model. I find that uncertainty shocks have adverse effects on the real economy and are deflationary in the short run. To recover the dynamic causal effects, the identification scheme is crucial.