Mar 15, 2017

Some Thoughts on a Geometric Mean Frontier

I know very, very little about geometric means and returns but I'm starting to think that I should know more.  I have tried to read a little bit about this but it can be a tough slog through some of this stuff.  That means that this post is not a demonstration of any kind of comprehensive knowledge. It is, rather, more of an attempt to convince myself that I know anything at all from my recent journey into geo land and an attempt to try to consolidate some of it "on paper" if I do.  As always, let me know if there are any major errors or omissions. 


Today, rather than tutorialize on geometric means in general -- it's a summary statistic that is a product of a series rather than a sum[4], it has some nice time-invariance properties in finance, it solves a problem with arithmetic returns being greater  over time than what really happens, it has a longish and sometimes controversial role to play in portfolio optimization, it's one of three classical means along with harmonic and arithmetic, it's a widely used measure of historical investment performance, etc. etc. -- which I can't do and others can do better, I am merely trying to generate a geometric mean frontier (GMF) from my own data and to ponder what it might mean to me, a normal retail systematic investor.  The basic idea, suggested by others (see, for example, Bernstein 1997), is to start with a standard Markowitz MV arithmetic frontier, convert the expected MV returns at a given risk level to geometric by some approximation[1], and then see what happens (or really to make a judgment about the idea of the proper "compensation for risk," I think).

I was too lazy to do anything other than grab a dataset I already had handy from a few months ago when I calculated for myself a MV frontier for random allocations of 5 ETFs over about 35 months somewhere in the 2014-2016 range.  That is not the best place to start for this particular task but at least I had a place to start.    I cheated a little because back then I had calculated geometric returns rather than arithmetic.  For this post I transformed them back to arithmetic by means described in the referenced articles. Let's pretend, though, that they were arithmetic from the start though calculating them from scratch would probably take less time than writing this paragraph about why I didn't.  The thing to do, then, I gather, is to transform, for the array of MV risk levels, the arithmetic returns to geometric by way of whatever of the many approximations seems appropriate[1].  In this case we will use Bernstein and Wilkinson's equation[2] for approximating a geometric return and ignore it's exact appropriateness or variance behavior from real historical returns:


In addition we will use, for comparison and what I'll loosely call "fun," Richard Michaud's equation 6a [3] for approximating geometric returns over some set of time intervals when the intervals are finite ( I use 2 periods below) rather than large or infinite:  


When one does this, the geometric mean frontier, relative to the traditional MV frontier, can look like this below…or at least it does for the data I hacked out for what I was trying to do earlier this year.  The lines probably look a little choppy because my sampling of the possible asset allocation choices was a little choppy.  That, and I couldn't figure out how to best transform the 5000 MV allocation sample-pairs of return & standard deviation that I used for 5 assets into one MV "line."  For today I just roughly binned the data, looked for the highest and least volatile points in the bin, and then did the best I could.  This is what I got.  Choppy. Close enough.   


So, the geometric returns along the geometric mean frontier -- as expected since geometric returns are always (correct me if I don't have this right but I think they are equal when variance is zero) below arithmetic -- are below the MV line and, this is important to look for in portfolio analysis, as risk rises to the right the curves look like they start bending a bit lower in relative terms. The blue line is Bernstein equation. The red line is the Michaud equation for geometric returns when the number of N periods is finite (2 periods in this case).  Red becomes more or less blue as N gets big.

The problem for me in this illustration, as lovely as it is, is that most of the charts of geometric frontiers I have seen lately in my reading seem to bend down a bit more than mine does. But then again the time period I happened to pick was relatively low vol and it's volatility (why don't these guys ever talk about spending…that's the real killer) that does funky stuff to expected (and realized) geometric returns.  In fact, I was so disappointed that I did not get the "bend," or what Michaud calls "critical points," that I decided I had to create some fake data to get it and so to complete this post.  The "critical point," according to Michaud, is an optimum along the curve of the GMF (not always present or maybe it's beyond the right end of the MV frontier) where the return is the highest and after which one might start to get lower (geo) return for higher risk, something that no one I know wants.  While you can kind of see it start to bend at 11-12% volatility, and I could have maybe imported charts from others websites as well to illustrate this concept, I wanted to exaggerate the bend in this post with my fake data to make a point.  Here, for example, is a fake MV frontier made with two assets.  The return/standard_deviation for asset 1 and 2 are 3%/7% and 9%/20% respectively. The correlation coeff is -.5.  I don't think this setup is wildly unrealistic but it is fake and the lower risk asset is maybe a little too high risk but it works for my point. Given those assumptions, then, the arithmetic and geometric mean frontier and the MV capital market line might look a little like this (ignore debate about risk-free rate levels while we do this since this is hypothetical):  


MV(A) is the standard Markowitz linear frontier. The tangency point is right about 40% allocated to the higher risk asset.  MV(G) is the Bernstein-converted (from above) geometric frontier given the parameters I set up.  The MV(G) "high vol" is an alternative completely fake setup with the same assets and returns and covariance as MV(G) except that asset 2 now has a 35% standard deviation rather than 20%.  I hope I got the math right (my chart looks a little extreme to me but who knows).  What it does show, whether I got it right or not, is that there is now a "critical point" somewhere around the 15% risk level (~75% allocated to the risk asset at that point) which is what I wanted and most articles on this subject try to show (usually at higher risk levels?).  The critical point (pun intended) here, and the end of this post, is that the lesson might be that sometimes when one thinks one is signing on for a single-period-mean-variance-optimal-or-maybe-riskier-but-higher-return portfolio, it is entirely possible that one might not be as efficiently well compensated for taking the additional risk [5] as one thinks when looking at it in geometric return terms…the returns that are the real true long term "effective" ones that one really does receive in life.  If that is all I learned from this (if I got it right) I guess it was worth the effort.


Notes ---------------------------------------

[1] there are quite a few approximations in quite a few papers.  Their derivation is beyond me. The approximations all seem to have both pros and cons as well as a ton of similarities and relationships between themselves.  This post is not a survey, though.  Mindlin, in the "Links" below has a good math-heavy survey of this topic.  There is, of course a lot of cross over and equivalency between the discussions of Mindlin, Michaud, Bernstein etc.

[2] http://www.efficientfrontier.com/ef/198/gmf.pdf

[3] from "A Practical Framework For Portfolio Choice"   Note that in the end he gives preference to the following equation which more or less converges with Bernstein's equation and Michauds 1/N formula as N gets large:

but see note 1 and Mindlin for a broader discussion of geometric approximations. 

[4] fwiw, the "textbook" definition of geometric average return:


[5] the "critical points" may or may not exist at all in any given portfolio analysis or may be beyond the relevant MV universe in question. I guess you gotta look.  It seems like someone with a portfolio made up of unusually volatile components who also has a long planning horizon is a good candidate for thinking this kind of thing through.  


Links ------------------------------------

Diversification, Rebalancing, And The Geometric Mean Frontier, William J. Bernstein and David Wilkinson 1997.  

A Practical Framework For Portfolio Choice, Richard Michaud, Journal of Investment Management 2003.  

On the Relationship between Arithmetic and Geometric Returns, Dimitry Mindlin, CDI Advisors 2011. 


Geometric Mean, Wikipedia 




No comments:

Post a Comment