I am in academia and begin to work on topics including portfolio optimization. I just read lots of paper discussing different extensions to the Markowitz approach, given different (possibly unrealistic) assumptions. Since it is an old method and I am away from the real finance organization, some facts from the real work might help our research a lot. I am just curious about:

  1. Is Markowitz mean-variance portfolio still widely used in companies?
  2. If yes to 1, it is well-known that Markowitz is prone to noise, for example, will large assets pools. Do you care about or really use those fancy method in papers?
  3. Many recent papers just consider the simplest mean-variance problem, i.e., without constraints or transaction costs. Is this practical? For example, in reality, do we need to first specify that, the exposure to certain a factor/sector should be among a range? Or what's the most commonly-required constraints, are they linear equality/inequality/ non-convex? Do they come from the government/investor, or simply the analyst's belief that they are helpful or easy to communicate with shareholders?

We academia researchers know some fancy tools, but I really want to stay close to the real needs. Thank you!

  • $\begingroup$ Hi: I don't know if this will help you but most of the academic research on markowitz theory focuses on the construction of more robust covariance matrices which are less sensitive to particular elements. This is also an interest in practice so it has good overlap. $\endgroup$
    – mark leeds
    Commented Jan 14, 2020 at 23:36
  • $\begingroup$ @markleeds Thank you! Yes, I am also working on similar topics. I am just unsure whether or not people really use Markowitz. If they do, then this kind of research will be meaningful. I am just afraid this line of research is only because of theoretical beautiful (mean and variance, easy to prove things). $\endgroup$
    – RunStat
    Commented Jan 15, 2020 at 21:05
  • $\begingroup$ My experience is that it is used but with different bells and whistles included in order to deal with reality. For example, in index fund management, you can put minimum bounds on the weights of each stock in order to force certain stocks to be held. Note that the journal of portfolio management might be useful for seeing how academic topics are put into practice. There are probably others as well but that's the one that stands out in my mind. $\endgroup$
    – mark leeds
    Commented Jan 16, 2020 at 22:13
  • $\begingroup$ That's a good suggestion! I will have a look at that journal. $\endgroup$
    – RunStat
    Commented Jan 17, 2020 at 0:55
  • $\begingroup$ Don't have enough to say to really warrant an answer, but in short, answer to Q1 is 'not really'. People do it, but as you noted it's widely understood to be 'flawed'. Related, Black-Litterman is probably used more but while better in some ways still suffers from similar issues. Depending on context, but I can't imagine many using MV absent some set of constraints--non-negativity of weights would probably commonly be the first constraint applied. $\endgroup$
    – Chris
    Commented Feb 17, 2020 at 1:57

1 Answer 1


Let’s put it this way. Classic MV is still used, but its shortcomings are universally appreciated. In its favour, the process is logical, conceptually intuitive, and non-quants easily understand it.

But the optimisation produces some very unintuitive results, no different to multicollinearity effects in regression analysis. That’s a harder one for the non-quants to grasp, especially if they have convinced themselves that the return on this is X and that is Y. If you really believe that Shell is +10% and BP is -10%, these optimal portfolio really is long/short 10x capital! The model is just a function of its assumptions, after all.

The way that most investment houses deal with this problem is akin to sampling errors in academia. They don’t wish to pick fights with their experts (who just might be right even if they are extreme!) So they oversample, requiring every forecaster to also star their assumptions on multiple related proxies.

Given these, the risk/asset-allocation group can infer outlier forecasts, and deduce more consensus forecasts for X given a multiplicity of forecasts for Y and Z. Ditto for Y and Z, of course.

They can also require their forecasters to produce a range around their forecasters (that the central portfolio team can freely adjust); that allows them to Monte Carlo the impact of different assumptions. This does not eliminate the Markowitz “noise”; but it does measure it, which informs how the central powers the choose to review and revise their assumptions to ensure more internal consistency with external intuition.

Outside of “risk parity” (in multi-asset) and “factor-based” “smart beta” (which demeans stockpicking gurus), Markowitz remains the default model. Academia tries to innovate the model. Finance tries to bootstrap its way around the problems, with workarounds within workarounds within workarounds. Most of these simply just trying to dial down the overconfidence and inconsistency in forecast that causes Markowitz and optimisation to habitually over-position portfolios.

Hope this helps, and happy to answer any follow-ups. No longer in this game professionally, after 20 years!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.