Forecasting financial and macroeconomic variables in the presence of instabilities and asymmetries

This thesis investigates several questions related to forecasting financial and macroeconomic time-series, aiming to address topical issues relevant for both academic and policy-making purposes. The first two chapters (Chapter 2 & Chapter 3) examine some of the main obstacles to exchange rate models’ predictive ability, with the presence of instabilities constituting a major challenge. Building on prior research suggesting that yield curves reflect expectations of market participants about future economic activity, in the first two chapters I study the empirical relevance of information contained in the term structure of interest rates in determining currency movements. Specifically, I conduct a comprehensive in-sample and out-of-sample evaluation where I investigate the importance of different sources of instabilities for delivering forecast improvements. Particular emphasis is placed on how the relationship between yield curves and exchange rate movements has evolved in the aftermath of the 2007/08 Global Financial Crisis that led to many central banks cutting policy rates to levels close to zero and adopting a variety of unconventional monetary policy measures. The behaviour of exchange rate fluctuations is, however, not the only one that is frequently characterized by features such as instability and deviations from Gaussianity. Precisely, important macroeconomic indicators exhibit similar characteristics. Evidence from the forecasting literature suggests that several predictors of, for example, output growth and inflation, that have substantial predictive content in one period, fail to maintain their usefulness in subsequent periods (see Stock and W Watson (2003) for a comprehensive discussion). That is, their predictive content is unstable over time (Rossi (2013a)). Chapter 4 focuses on real-time forecasts of GDP growth, inflation, short-term interest rates and unemployment. This work is concerned with the problem of incorporating external forecasts from survey participants into medium-term macroeconomic forecasts from conventional unrestricted vector autoregressive (VAR) specifications. I contribute to the macroeconomic forecasting literature by proposing a methodology that captures the documented non-Gaussian behavior of the macroeconomic time series under investigation. Chapter 2 reviews the in-sample relationship between relative yield curve factors and exchange rate movements considered by Chen and Tsang (2013). I extend their empirical analysis to a wider set of currency-pairs (G10) and for a longer sample that includes the 2007/08 Global Financial Crisis. In addition, moving beyond in-sample considerations, I examine the relevance of the term structure of interest rates in generating accurate exchange rate point forecasts. My analysis provides evidence on how the interaction between yield curves and exchange rate fluctuations has substantially changed following the onset of the Great Financial Crisis. Given this widespread evidence of instabilities, I utilize forecast averaging methods in the presence of a single break of unknown timing. These techniques entail averaging forecasts from the same model, but over different estimation windows. I show that properly accounting for the uncertainties associated with the size of the estimation window and the unknown timing of structural breaks should be critical to the choice of modeling method. In the second essay (Chapter 3), I examine the relevance of Nelson-Siegel yield factors in generating density exchange rate forecasts within a framework that allows for the presence of structural instabilities. I find that controlling for the relative level, slope and curvature within a setting that also allows for time-evolving first and second-order moments delivers accurate predictive densities. In particular, my analysis shows that the information content of yield-curve differentials offers consistent density forecast accuracy gains relative to the naive random walk and a multivariate specification that jointly models all currency-pairs. Moreover, introducing time-varying slope and volatility parameters results in improvements in density forecasting performance relative to a term-structure specification that imposes the assumption of time-invariant parameters. A decomposition that uncovers the benefits associated with allowing time-variation to enter either the slope or the volatility parameters suggests that the latter is responsible for gains at both short- and long-term horizons. In further analysis I illustrate that my results are robust to alternative benchmark specifications that employ widely-used exchange rate predictors (e.g. inflation differentials) and explicitly account for the persistently low interest rate environment of the last decades. Moreover, I show that utilizing information from yield-curve differentials within a time-varying framework generates economic value in a mean-variance asset allocation setting. A second area where researchers have documented an unstable behavior is macroeconomic forecasting. Vector autoregressions (VARs) with a variety of general error distributions are useful for modeling macroeconomic time series and have been found to be successful in medium-term forecasting. However, their short-horizon predictive ability has been rather limited. To address this limitation, recent work has proposed incorporating external information extracted from a survey of professional forecasters into real-time macroeconomic predictions of a Bayesian VAR. The method of entropic tilting achieves this by modifying the baseline VAR distribution such that it matches moment conditions of interest. Existing papers make use of survey densities that have been aggregated using an equal-weight combination scheme and restrict their attention to the first- and second-order moments. As a result, the conventional entropic tilting forecasting methodology fails to capture asymmetry, a feature that many macroeconomic time series exhibit. In Chapter 4, I propose a modification to the standard relative entropy approach which allows for asymmetry in the macroeconomic variables. The proposed methodology involves tilting the VAR distribution towards the aggregate survey forecast moments that have been appropriately reshaped in order to match non-Gaussian features of the sample data. Transforming the aggregate projections entails the adoption of a non-parametric Empirical Cumulative Distribution Function (ECDF) fitted to the macroeconomic variables of interest. I illustrate this methodology with an application examining real-time xv forecasts for U.S. GDP growth, inflation, unemployment and the Federal Funds rate over an evaluation sample from 1988:3 to 2017:4. I consider a variety of VAR models, ranging from simple fixed-parameter and Gaussian-errors to time-varying volatility and non-Gaussian errors. Results across models indicate meaningful gains in terms of both point and density forecasting accuracy relative to individual multivariate specifications and existing tilting methods