On the Black–Litterman Model: Learning to Do Better

In this article, the authors study the performance of the Black–Litterman model (BLM) and compare it to the traditional mean–variance theory (MVT) of Markowitz (1952) and Sharpe (1964). They begin with the standard Bayesian learning on which the BLM is based (but the existing literature does not follow). Then, they perform a series of tests of the BLM using machine learning tools and view specifications consistent with the existing literature. Their empirical evidence (which uses 30 years of monthly data from January 1991 till December 2020) suggests that the BLM is highly sensitive to the specification of the view. Given that the view is arbitrary (even though in our article, they are rule based), it is quite a challenge to use the BLM in an actual situation. A great amount of caution must be exercised in specifying the view and its corresponding required return. This validates the previous result that BLM specification of views is very important and there is no consistent manner how one can specify a winning portfolio.