Feature selection for portfolio optimization

Thomas Trier Bjerring, Omri Ross, Alex Weissensteiner

    Research output: Contribution to journalJournal articleResearchpeer-review

    1060 Downloads (Pure)

    Abstract

    Most portfolio selection rules based on the sample mean and covariance matrix perform poorly out-of-sample. Moreover, there is a growing body of evidence that such optimization rules are not able to beat simple rules of thumb, such as 1/N. Parameter uncertainty has been identified as one major reason for these findings. A strand of literature addresses this problem by improving the parameter estimation and/or by relying on more robust portfolio selection methods. Independent of the chosen portfolio selection rule, we propose using feature selection first in order to reduce the asset menu. While most of the diversification benefits are preserved, the parameter estimation problem is alleviated. We conduct out-of-sample back-tests to show that in most cases different well-established portfolio selection rules applied on the reduced asset universe are able to improve alpha relative to different prominent factor models.
    Original languageEnglish
    JournalAnnals of Operations Research
    Volume256
    Issue number1
    Pages (from-to)21-40
    ISSN0254-5330
    DOIs
    Publication statusPublished - 2017

    Keywords

    • Agglomerative hierarchical clustering
    • Feature selection
    • Parameter uncertainty
    • Portfolio optimization

    Fingerprint

    Dive into the research topics of 'Feature selection for portfolio optimization'. Together they form a unique fingerprint.

    Cite this