Juan Arismendi-Zambrano

We use and test different machine learning methods suited for high-dimensionality data, to test five-minute-ahead return forecasts of Brazilian stock returns from the the stock index (Ibovespa). The methods used are Ridge Regression, LASSO, elastic net, PCR and PLS. We use as explanatory variables the lagged returns of all the stocks comprising the stock index of Brazil (Ibovespa). To assess the out-of-sample forecast performance of the different machine learning methods in this emerging market setting, different portfolios were formed, such as one portfolio selecting a subset of the most volatile stocks, or a portfolio that exploits the predictability of machine learning methods. The results show that, when accounting for transaction costs, the Ridge Regression outperforms the other methods and can provide profitable trading strategies, in particular for the portfolio with the most volatile stocks, indicating a potential predictability of non-linear effects of clustered volatility.

Even though a multi-factor linear asset pricing model can be equivalently represented in a Beta or in a stochastic discount factor (SDF) form, its inferential efficiency and pricing accuracy features may differ when estimated by the generalized method of moments (GMM), both in small and in large samples. We test the estimated variance of the equivalent Beta and the SDF parameters under GMM with small sample bootstrapped simulations and derive the analytical asymptotic variance and find that the SDF approach is likely to be less efficient but to yield more accurate pricing than the Beta method. We show that the main drivers of this trade-off are the higher-order moments of the factors that play an important role in the estimation process.

We estimate the personal communication risk-premium profile of the U.S. Federal Reserve(Fed) Chair by measuring a new dataset of the sentiment revealed by their public statements during their tenure. We analyze the impact of such Fed communications’ sentiment risk on the uncertainty of the monetary policy, and the market price discovery process of interest rates, in the aftermath of the Federal Open Market Committee (FOMC) meetings. After controlling for the evolving state of the economy surrounding the meetings, we find that there is a significant statistical and economic difference in the communications’ sentiment that is heterogeneous across Chairs, depending on their personal traits. The sentiment in the Chairs’ communications plays a role in moderating the potential surprises in the Fed announcements, and it can be effectively used as a tool for controlling and measuring monetary policy shocks.

This paper investigates the counterparty credit risk of interest rate swaps positions using the credit valuation adjustment (CVA) measure, and examines the potential dependency relationships between the probability of default (PD) and exposure at default (EAD). We empirically tested, using interest rate swaption implied market volatilities, three tail dependency models: a Basel III Committee independent model, a Gaussian copula dependent model, and a Wrong Way Risk (WWR) with copula dependency approach. The results show that the CVA underestimation when using a Gaussian copula for modelling the dependence of PD and EAD is about 51%–362% compared to using WWR, and the underestimation between using the standardised Basel independent model and using the Gaussian copula is about 527%–1609%, including the period of the 2007/2008 crisis. This has important implications for regulators, financial institutions, and credit risk managers when calculating counterparty risk.

The price of a European option can be computed as the expected value of the payoff function under the risk-neutral measure. For American options and path-dependent options in general, this principle cannot be applied. In this paper, we derive a model-free analytical formula for the implied risk-neutral density based on the implied moments of the implicit European contract under which the expected value will be the price of the equivalent payoff with the American exercise condition. The risk-neutral density is semi-parametric as it is the result of applying the multivariate generalized Edgeworth expansion, where the moments of the American density are obtained by a reverse engineering application of the least-squares method. The theory of multivariate truncated moments is employed for approximating the option price, with important consequences for the hedging of variance, skewness and kurtosis swaps.