Search

This thesis evaluates and proposes forecasting methods for two primary categories of investments, stocks and bonds, aiming to advance our understanding of the underlying risk{return attributes and to explore the scope for enhanced predictive content in practice. More specifically, the thesis identifies and examines the cross-sectional and time-series determinants of stock returns and government bond yields from a predictive perspective. Chapter 2 examines the information content of variance swaps in predicting the cross section of stock returns. A non-parametric approach is employed to extract principal components as underlying predictive variables, which mitigates issues concerning model misspecification and estimation uncertainty with parametric approaches. The variance swap components exhibit substantial predictive power for the cross-sectional stock returns. The last few principal components, despite carrying little weight in terms of explaining the variation in the variance swap surface, contribute considerably towards predicting stock returns. The information value of variance swaps is attributable to their strong associations with financial and macroeconomic state variables that track investment opportunities, and with various moment measures that characterise the variations in aggregate stock market returns. Providing additional information beyond that afforded by benchmark factors, the variance swap components, combined with Fama{French{Carhart factors, deliver superior forecasts. Chapter 3 evaluates the predictability in government bond markets implied by yield forecasting approaches that exploit macro-finance interactions in a sample extended to the post-global financial crisis period, which spans 1970:M1{2016:M12. This chapter places particular focus on the macro-yields model featuring unspanning restrictions proposed by Coroneo, Giannone, and Modugno (2016), which has demonstrated satisfactory performance in forecasting the yield curve and excess bond returns within the 1970:M1{2008:M12 timespan. While the model's in-sample fit alters little, the out-of-sample predictability deteriorates substantially as a result of severe underprediction of all three Nelson-Siegel yield factors during the 2007{ 09 global financial crisis (GFC), and the overprediction of yields in the subsequent recovery. The chapter further identifies the leading causes of the deterioration and investigates the stability of yield dynamics and macro-financial linkages, by constructing conditional forecasts and counterfactual scenarios. Even after controlling ix for the anomalous behaviour of the front-end of the curve, the predictability of longer government bonds still suffers from substantial declines. Meanwhile, the instability of yield dynamics and macro-finance interdependence is revealed in the context of the linear-Gaussian dynamic factor representation. These results collectively suggest that the severe model degradation cannot be solely attributed to the constraints on the dynamics of the short-term rates. It also emphasises the importance of addressing the growing complexity in yield dynamics and macro-finance interactions after the global financial crisis, which cannot be adequately accommodated by the existing macro-finance term structure framework. Chapter 4 explores the potential of the copula framework to handle non- Gaussianity in the context of macro-finance term structure modelling and forecasting. The non-Gaussian macro-yields model proposed here accounts for the asymmetry and tailedness in yield distributions through non-parametric marginal densities, as well as explicitly addressing the cross-sectional and serial dependence via the state-space inversion copula, and in doing so retains a latent dynamic factor structure amenable to efficient implementation. Regardless of maturities and forecast horizons, exploiting the informational content of macroeconomic data in a non-Gaussian setting improves both in-sample and out-of-sample forecasting performance relative to the Gaussian macro-yields model over the 1970:M1{2016:M12 period. Furthermore, the non-Gaussian macro-yields model demonstrates overwhelming superiority in predicting excess bond returns over the expectations hypothesis and the prominent macro-financial predictors. It also compares favourably with the random walk in forecasting the yield curve over medium- to long-term horizons. Lastly, this copula-based approach affords a technically convenient and extensible means of accommodating high-dimensional macroeconomic datasets and the growing complexity of post-crisis yield movements, which facilitates further investigation into their practical implications for government bond modelling and forecasting.