Support vector decision making

The paper develops a support vector machine (SVM) for binary decision-making within a utility framework. Given an information set, a decision-maker first predicts a binary outcome and then selects a binary action based on this prediction to maximize expected utility, where the utility function can depend on the action taken, observable covariates, and the binary outcome subsequently realized. The proposed maximum utility SVM differs from the traditional SVM in four key aspects. First, as a conceptual innovation, it incorporates the optimal cutoff function as a separate and special covariate. Second, there is a sign restriction on this special covariate. Third, it accounts for the dependence of the utility-induced loss on both the covariates and the binary outcome. Finally, it allows the margin to differ across different classes of outcomes. The paper proves that the proposed method is Bayes-consistent under the maximum utility criterion and establishes a finite-sample generalization bound. A simulation study shows that the proposed method outperforms existing methods under the data-generating processes considered in the literature.