Abstract
The supervised portfolios approach is an effective asset allocation strategy that engineers optimal weights before feeding them to a supervised learning algorithm. Yet, supervised learning algorithms are often seen as opaque, which undermines trust in those models, thereby limiting their adoption. To alleviate this issue, the authors apply an enhanced version of RuleFit, an intrinsic interpretable algorithm, that transforms a black box nonlinear predictive algorithm into a simple combination of rules. It fits a sparse linear model that includes feature interactions derived from decision tree ensembles. The first empirical analysis illustrates that the interpretable approach is statistically as accurate as gradient boosting on three different investment universes. The second analysis highlights which characteristics and interactions matter for an equity portfolio manager.