Using Diverse Local Optima for Setting Kernel Parameters in Support Vector Regression: Forecasting Emerging Market Credit Spreads

Abstract

We propose a novel approach for determining support vector regression (SVR) kernel parameters in the presence of multiple local optima. In contrast to existing approaches focusing on identifying a single "best" tuning parameter setting, an impractical goal in many financial market applications, our framework employs a global optimization algorithm to produce a collection of competitive SVR kernel parameter candidates, and applies the Model Confidence Set test to select the most accurate set from the collection of promising candidates. We use our approach to predict credit spreads for four mature emerging market sovereign borrowers. Combining our forecasts into simple forecast combinations and contrasting them against random forest, standard SVR, conventional random walk, and linear model forecasts, we find that forecast combinations from our most accurate model sets deliver substantial gains in forecast accuracy. Furthermore, the analysis of the forecasting results provides useful insights into credit risk pricing by international investors.

Keywords: Support Vector Regression, Kernel parameters, Forecasting, Model Confidence Set, Emerging market credit spreads. JEL Classifications: G17

JEL Classification: G17, G15, C53, C51