Essays in Empirical Asset Pricing and International Finance

Abstract

Chapter 1 introduces the Supervised Dynamic Orthogonal Components (sDOC) method as a novel framework for forecasting the equity risk premium out-of-sample. sDOC advances traditional linear dimension-reduction techniques - most notably Principal Component Analysis (PCA) - by integrating machine learning-based feature selection with the construction of dynamic, volatility-sensitive orthogonal factors. Unlike PCA, which does not capture nonlinear relationships or volatility interactions among predictors, sDOC explicitly models these elements, thereby achieving superior forecasting accuracy. The method is applied to 63 U.S. monthly industry equity portfolios and evaluated against several benchmark models: PCA, univariate GARCH(1,1), random forest (RF), and a three-layer neural network (NN3). Empirical evidence shows that, across most industries, sDOC outperforms PCA, univariate GARCH(1,1), and RF, while delivering predictive performance comparable to NN3. This superiority is evident across multiple evaluation criteria, including out-of-sample R², Mincer-Zarnowitz R², gains in average excess portfolio returns, and a range of risk-adjusted metrics such as the Sharpe ratio, Sortino ratio, Calmar ratio, and gain-to-pain ratio. Moreover, the rolling out-of-sample R² underscores sDOC's adaptability under both calm and turbulent market conditions, positioning it as a robust and interpretable tool for asset return prediction.

Chapter 2 proposes a modeling strategy to investigate the transmission of shocks - whether through interdependence, contagion, or decoupling - between two asset markets during periods of heightened volatility. The regime-switching model captures co-movements in both the mean and volatility processes of asset returns. The mean dynamics incorporate PCA-derived factors that reflect global and market-specific influences, while the variance-covariance structure accounts for common and idiosyncratic shocks, each governed by an independent Markov-switching process. Contagion or decoupling is defined as occurring when a high-volatility idiosyncratic shock originating in one market significantly alters the interdependence of mean returns across the considered assets. To statistically detect these phenomena, we employ a novel bootstrap-based Student's t-test procedure. Applying this methodology to sovereign bond market pairs across three Latin American countries, we find evidence that, on average, decoupling has occurred in the Brazil-Mexico and Argentina-Mexico pairs. For the remaining combinations, our results suggest that shock transmission is characterized primarily by interdependence.

Chapter 3 investigates the impact of country-specific major export commodity shocks on domestic equity returns. Employing a Markov-switching framework, we jointly model and estimate commodity-equity return pairs, capturing co-movements in both the mean and the variance-covariance structure of these assets. Co-movement in the mean reflects the influence of global risk, while the variance-covariance structure accounts for common and idiosyncratic shocks, each governed by an independent Markov-switching process. We then examine the transmission of shocks during crises, focusing on cases where shocks originate in a country's commodity market (idiosyncratic shocks) and spill over into its equity market. Contagion is defined as the amplification of idiosyncratic commodity shocks within equity market returns during a crisis. Using data from six major commodity-exporting countries between 2006 and 2024, and applying likelihood-based tests, we find evidence of contagion during the Global Financial Crisis - most notably in the oil-equity linkages of Norway and the United States. Out-of-sample forecasts further demonstrate that our proposed model consistently outperforms several widely used benchmarks, particularly in predicting equity returns.