Pest Control: Eliminating Nuisance Allocations through Empirical Asset Class Identification

There is a seemingly infinite number of ways to partition the investment universe into asset categories. Too fine a partition can lead to very small allocations to certain categories in asset allocation models. When populating those categories with actual strategies, costs increase because due diligence and reporting are not costless activities. The right number of asset categories for building allocations should depend on balancing the marginal benefits and marginal costs of a finer partitioning of the investment universe. In this article, the authors use different machine learning techniques to help quantify these trade-offs. Two unsupervised learning techniques—exploratory factor analysis and hierarchical cluster analysis—are used to identify asset classes. A supervised learning technique—a regression tree—then is used to identify the most important basis for US equities, a specific asset class identified by the unsupervised learning techniques. TOPICS: Big data/machine learning, security analysis and valuation, performance measurement