This dissertation studies the forward premium puzzle (FPP) and short-term exchange rate forecasting.
Chapter 1 studies the empirical behavior of the FPP over different subsamples instead of an average effect for the whole sample period as what is typically done in the literature. We find that the estimated slope coefficients from the Fama regression vary considerably from period to period. The signs of the slope estimate could be both significantly positive and negative. Our contribution is to show that the variation of the slope estimates is not random, rather it is driven by a common factor. We document a link between the variation and investors' long-run uncertainty about the economy. The long-run uncertainty index is specific to individual countries and defined as either a large fall in the real GDP growth rate or an inflation hike compared to past levels. We find that the long-run uncertainty index and its lags contribute to the positiveness of the slope estimate. The effect lasts longer for developed countries than emerging ones. The FPP exists if there is no long-run uncertainty about the economy but disappears with such uncertainty.
Chapter 2 provides a potential theoretical framework to understand the empirical facts described in Chapter 1 based on Li and Tornell (2015). They show that the robustness against model misspecification can generate both positive and negative Fama slope coefficients, depending on investors' beliefs about the relative importance of transitory and persistent interest rate shocks. But they miss one step linking the economic fundamentals to the assumed interest rate differential model. We fill the gap using the long-run risk model with two variables: real consumption growth and inflation. We map the persistent interest rate shocks to long-run shocks to either consumption growth or inflation, which matches the long-run uncertainty defined in Chapter 1. We then qualitatively explain the empirical facts of time-varying slope estimates.
Chapter 3 implements an empirical forecasting strategy based on what the Federal Open Market Committee (FOMC) says after their regular meetings. We use several techniques from natural language processing including bag-of-words, latent semantic analysis and vector space model to construct nontraditional predictors from three types of text documents released by the FOMC. We apply a machine learning algorithm called support vector machine to forecast individual G10 currencies and also build a portfolio of all G10 currencies. For the portfolio, our out-of-sample forecasts have success ratios more than 50\% for short-term prediction (less than 6 weeks) except for the 1-month horizon. Our best performance can be found for 1-week forecasting horizon. Eight out of nine currencies, as well as the portfolio, can beat the random walk model significantly using the weighted directional test.
