Some preliminary qualitative observations

What lessons can the prudent portfolio manager draw from the sensitivity analysis presented in the previous chapter?

The strong dependence of the recommended allocations on the most-difficult-to-ascertain set of quantities, the expected returns, may, at first blush, appear somewhat dispiriting. Indeed, after running our sensitivity analysis with behaviourally plausible coefficients of risk aversion, one can justifiably conclude that it would be foolhardy to rely on even the most careful construction of a Bayesian net to ‘read off’ with confidence a single set of allocations from a graph such as the one in Figure 26.11: yes, we may well read a precise set of weights today, but small, unavoidable changes in the input expected returns tomorrow (changes which have nothing directly to do with the Bayesian-net method per se) could give rise to very different allocations. How can we get around the problem of the instability of the optimal allocations to the various asset classes?

To answer this question, we stress again that the sensitivity of the allocations to small changes in expected returns is not an artifact, or a peculiar feature, of the Bayesian-net approach. The analysis presented in the previous chapter (but see also in this respect Chapter 8) shows that it is an unavoidable feature of all allocation methods based on mean-variance optimization.