City Research Online

Abstract

In the first essay, I expand on the work of Borghi et al. (2018) by comparing the in-sample performance of their proposed ML estimator of time-varying sensitivities, which draws from Mikkelsen, Hillebrand, and Urga (2019), against the alternative rolling least square estimator. The main finding of my analysis is that the rolling OLS estimator is characterised by a great degree of instability, which is driven by the interplay between window size and sampling frequency. If one selects the window size arbitrarily the instability in the estimates can be pronounced, and the benefits of fitting a dynamic model as opposed to an equivalent static-loadings representation become slim. Overall, the ML estimator of Borghi et al. (2018) dominates the rolling OLS under many aspects.
In the second essay, I expand on the work presented in the first essay in two ways. I com pare the explanatory power of the three-factor model of Borghi et al. (2018), that features a combination of observed and latent factors, against more traditional factors constructed from firm attributes, and I evaluate the out-of-sample predictive performance of rolling OLS betas in forecasting future return patterns, accounting explicitly for the window selection problem. The main finding of my analysis points to a dual role of the rolling least square estimator when I employ the measures in Kelly, Palhares, and Pruitt, 2021 to gauge the models performance. While the short-window approach provides the best results in a contemporaneous-equation setting, for predictive purposes including too little observations for estimation causes the betas to be noisy, which in turn results in forecasts with little predictive power. Out-of-sample, the trade-off between the length of the window and the variance of the estimator is resolved around the two-year point, and this is true across all model specifications. I find that the choice of the window length alone accounts for about ±10% of the factor model’s out-of-sample fore casting performance. Analysing the out-of-sample relative performance of alternative model specifications that include observed factors, I find the performance of Fama French factors deteriorates significantly with respect to the in-sample analysis.

The third essay turns to the analysis of factor premia in international sovereign bonds. I identify the factors with country-style characteristic-based portfolios such as momentum, value, and low-risk, and study their performance under two dimensions simultaneously, issuer and maturity-wise. My analysis reveals substantial variation in the factor premia across the cross-sections, which does not support the view of Asness, Moskowitz, and Pedersen (2013) and Frazzini and Pedersen (2014) on their unifying pricing ability across countries and asset classes. I find that risk-adjusted returns are decreasing in the maturity of the bonds for momentum strategies. When analysed across countries, momentum produces consistent statistically significant Sharpe ratios, however this is not true for value and low-risk. The former shows low and insignificant returns across countries, while low-risk yields statistically significant premia only for Euro Area bonds. Contrarily to what reported in previous literature, I find no supporting evidence for the existence of risk premia for characteristics-based global portfolios.

Publication Type:	Thesis (Doctoral)
Subjects:	H Social Sciences > HG Finance
Departments:	Bayes Business School > Bayes Business School Doctoral Theses Bayes Business School > Finance Doctoral Theses

Export

Downloads