Benchmark Independence and Portable Alpha in Active Portfolio Optimization
Summary
The document presents a mean-variance framework for a fully invested portfolio, then decomposes holdings into benchmark weights and active weights. Under the stated setup, the active-weight optimization maximizes expected active return while penalizing active variance, subject to active weights summing to zero. The resulting expression contains expected returns, the covariance matrix, and risk tolerance, but no benchmark weights, motivating a question about when active weights and their risk and return can be considered benchmark-independent.
It then asks how portable alpha works in practice, especially when benchmarks contain different securities and therefore have different dimensions. The discussion provides the optimization equations and a conceptual example contrasting an index-relative equity portfolio with a market-neutral fund, but it does not supply an answer or a mapping method between benchmarks. The independence claim is therefore tied to the stated model and constraints; implementation across different universes requires additional assumptions about eligible assets, exposures, and risk.
Key ideas
- The active portfolio is represented as benchmark holdings plus active deviations whose weights sum to zero.
- In the stated mean-variance model, the optimized active weights are expressed without benchmark weights.
- Portable alpha refers to applying a return-seeking strategy alongside a different benchmark exposure.
- Different benchmark universes have different dimensions, and the document leaves the practical mapping unresolved.
Tags
Full text
# "Porting" an alpha strategy to a different benchmark
# "Porting" an alpha strategy to a different benchmark
I'm reading about the mean-variance optimization of active portfolios. A bit of prior background from the book I'm reading: the author discusses the mean-variance optimal portfolios without cash, which amounts to solving the following optimization problem:
Maximize: $w^Tf - \frac{1}{2}\lambda (w^T\Sigma w)$, subject to $w^Ti = 1$,
where $w$ and $f$ are column vectors of the weights and returns (respectively) of all securities, $\Sigma$ is the covariance matrix, $\lambda$ the risk tolerance and $i = (1, ..., 1)^T$ is just a column vector of all $1$'s. The optimal weight vector turns out to be:
$$w^* = \frac{\Sigma^{-1}i}{i^T \Sigma^{-1}i} + \frac{1}{\lambda}\frac{(i^T \Sigma^{-1}i)\Sigma^{-1}f\ -\ (i^T \Sigma^{-1}f)\Sigma^{-1}i}{i^T \Sigma^{-1}i}$$
Next, while considering an active portfolio, we can decompose the portfolio into benchmark and active weights - $w = b+a$. Since $w^T i = 1$ and $b^T i = 1$, $a^Ti = 0$. So the optimization problem in this case is:
Maximize: $a^Tf - \frac{1}{2}\lambda (a^T\Sigma a)$, subject to $a^Ti = 0$. The optimal active weight vector is:
$$a^* = \frac{1}{\lambda}\frac{(i^T \Sigma^{-1}i)\Sigma^{-1}f\ -\ (i^T \Sigma^{-1}f)\Sigma^{-1}i}{i^T \Sigma^{-1}i}$$.
So far so good. Now the interpretation given in the book is as follows:
> ...it (the active weights) is independent of the benchmark. Consequently, the expected active return or alpha and the active risk are also independent of the benchmark.
I can't understand how this claim follows from the equations above. Secondly,
> It is therefore theoretically feasible to utilize or port it on any benchmark. In other words, two active equity portfolios managed against two different equity benchmarks could have the same active weights. For instance, the active weights of an equity portfolio managed against S&P 500 index could be the same as the weights of a long-short market-neutral hedge fund. This is the idea behind the so-called portable alpha strategies, i.e., the alpha or excess return generated from a strategy can be ported onto another different benchmark.
Could someone please explain what is meant by "porting" on to a benchmark? How can you port an $n$-dimensional active weight vector meant for an index with $n$ stocks on to another benchmark with $m \neq n$ stocks?Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.