improvements in the precision of the estimate. Intuitively this makes sense as relative views do
not provide an improved estimate of the mean, just extra information on the relationship between
the estimates. We can measure the precision of the estimates by summing the unconstrained
weights. We can compute an effective posterior measure of uncertainty/precision as shown below
(16) =
1−∑
1
n
wi
∑
1
n
wi
The value η in formula (16) can be compared to τ and can be used to measure the uncertainty in
the prior or posterior estimates of the mean. When viewing the prior estimates, η = τ. We can
compare η between the prior and the posterior to determine the relative improvement in the
precision of the estimates.
One of the interestingly artifacts of the Black-Litterman model is that while the estimated return
of an asset without views can change, the unconstrained weight of the asset in the portfolio does
not change at all. We can prove this point by examining the formula 17 in He and Litterman
(1999). Λ is a k x 1 matrix with one row per view. The PTΛ term will always be zero for an asset
in no views. Under the Canonical Reference model, our investor with less than 100% confidence
in their prior estimates is not 100% invested, but is only invested in the fraction 1/(1 + τ). This is
because of formula (6) which shows the prior dispersion of realized returns about the estimated
mean is (1 + τ)Σμ. The asset weights in an unconstrained portfolio based only on the prior will be
weq/(1 + τ). Because the posterior precision of the estimated mean will be equal or higher, the
investor will invest an equal or larger fraction of their wealth in the portfolio and the asset
allocation will experience some change solely because of this change. Of course, since most
investors will be using constrained portfolio optimization of some sort, the final asset weights
will most likely change even when an investor has no view on a specific asset.
Note that we will use unconstrained mean variance optimization to illustrate the results of the
model, but it in no way implies a requirement to use either constrained or unconstrained mean
variance optimization with the Black-Litterman model.
Now we will work an example using both the Canonical Reference model and the Alternative
Reference model. The details of the example can be found in Appendix A. We examine two
scenarios
• Relative Views – Investor has two relative views
• Absolute View – Investor adds an absolute view on Germany with return = return from
the first scenario.
This construction should allow scenario two to illustrate the difference in the unconstrained
weights caused solely by the updated posterior covariance matrix.
Table 1 shows that when using the Canonical Reference Model the unconstrained weights of the
assets without views (Japan) do not change from the prior unconstrained portfolio. Note that
because the investors confidence in the prior is less than 100%, the prior asset weights differ
from the equilibrium by a factor of 1/(1 + τ). We can also see that an absolute view which only