Resumen de Essays in economic forecasting
This thesis consists of three self-containing chapters, all of which are related to economic forecasting and contain both, theoretical and empirical results. Chapter one and three present and evaluate models for multivariate and univariate density forecasts. Chapter two provides a framework to test for time variation in the forecasting performance of competing models.
In chapter 1, I present a methodology to estimate multivariate density forecasts based on univariate densities from survey data. Survey-based predictions are often competitive to time series models in terms of their forecasting performance but have a univariate focus. My methodology exploits the information in the surveys' marginal densities for the estimation of the multivariate densities. I demonstrate the importance of the multivariate aspect for new measures of the state of the economy and a novel measure of joint macroeconomic uncertainty. A stronger distributional dependence between the variables has different implications for the two types of measures. It tends to increase the probability of \say{recession-type} events and reduces uncertainty. Empirical results based on SPF data from the euro area and the U.S. show that the survey-based joint density forecasts are competitive to current benchmark econometric models. When considering joint macroeconomic uncertainty, the dependence has sizeable effects on my uncertainty measure in the aftermath of the Great Recession, a feature of the data that existing measures would not capture.
Chapter 2 proposes novel tests for forecast rationality, which are robust under the presence of Markov switching. Existing tests focus on constant out-of-sample performances or use non-parametric techniques; consequently, they may lack power against the alternative of discrete switches. Monte Carlo results suggest that the tests we propose have better power than existing tests in detecting Markov switching deviations from unbiasedness or efficiency. In an empirical investigation of the forecast rationality of the Blue Chip Financial Forecasts for the federal funds target rate, we find evidence against forecast unbiasedness. During periods of monetary easing, the forecasters tend to overestimate the future interest rate systematically. The size of the systematic bias component is around 25 basis points, a typical interest rate move of the federal reserve.
In chapter 3, I provide an empirical investigation of the real-time forecasting performance of quantile regressions for predicting different vintages of real GDP growth. Given a large number of potential predictors, I pool univariate models using an equal-weighting scheme and bayesian model averaging. My results indicate that equal-weighting outperforms bayesian model averaging and that forecasting first-release GDP growth is best done by also basing the in-sample estimation on first-release data only. When the predictive performance is compared to stochastic volatility models, the pooled bayesian quantile regressions outperform the competitor models at the one-quarter-ahead forecast, in particular for quantiles above the median. For horizons of four quarters-ahead, the results speak in favor of quantile regressions but exhibit substantial time-variation - the multivariate stochastic volatility models tend to perform better during recession periods but worse after the great recession.
