Skip to content

random portfolios

13 messages · Kevin Dhingra, Brian G. Peterson, Ross Bennett +3 more

#
Hello everybody,

I have been using the random_portfolios function from the
`PortfolioAnalytics` package to simulate the range of possibilities for
return paths at each step under various portfolio constraints / mandates
for evaluating mutual fund managers. As more managers are added to the
universe, however, and more simulations are needed, the pure R
implementations get pretty heavy and hard to scale. I was wondering if
there has been any work out there thus far on implementing any of the three
random portfolio generation methods (sample, simplex, and grid search) at a
lower level, using something like `Rcpp` to enhance the efficiency of these
algorithms?


Any help/feedback is much appreciated.

Thank you,
#
On Mon, 2017-03-20 at 15:09 -0400, Kevin Dhingra wrote:
We've discussed it, but I can't say that it is terribly high on our
list of priorities. ?

In most cases, no more than 1-2k portfolios should be required to get a
fair view of the feasible space given your constraints and objectives.

We'd be happy to work with you if you want to craft a patch to use C or
Rcpp for this.

Regards,

Brian
#
Brian,

Thank you for a quick reply. I will soon be working on that problem and
from what I have played with so far, it is unlikely that for our example
~2k portfolios will be enough (really hoping it would) to get a good sense
of the feasible space and seems like I need to implement an Rcpp version of
the random portfolios function. I will be happy to collaborate and share my
code once i get a decent handle on it locally for the purposes of our
current project.

Regards,
Kshitij Dhingra



On Mon, Mar 20, 2017 at 3:17 PM, Brian G. Peterson <brian at braverock.com>
wrote:

  
    
#
Kevin,

Can you give us a sense of the number of assets in the portfolio and
the constraints? That will help us understand where the potential
bottlenecks are in the random portfolio generation. For example,
generating a set of random portfolios for box and weight constraints
if relatively fast, but adding group or position limit constraints
makes the algorithm more complicated and slower.

Thanks,
Ross


On Mon, Mar 20, 2017 at 2:35 PM, Kevin Dhingra
<kevin.dhingra at appliedacademics.com> wrote:
#
Hi Ross,

Sure. Even though I have not profiled the bottlenecks quite in detail as of
yet, i will give you a decent idea of the problem I am working with. I can
have multiple indices with as much as 2000 assets with group, position and
turnover limits (Not sure if i can increase the speed by removing
constraints and doing rejection sampling later). In order to generate a
daily possible set for the market in this case, I was playing around with
~4-5 thousand permutations. Also I think I will end up using the "sample"
method because of the type of constraints we have and as you already have
mentioned that method is the slowest (takes about 30 times the time using
"simplex" for the same constraints). Adding box and position limit
constraints are causing it to run a bit slower (but its not a big
difference). I can always provide a more thorough analysis of the potential
bottlenecks with a lot more detail when I have a chance to start working on
translating it to cpp

Thank you,

On Mon, Mar 20, 2017 at 4:04 PM, Ross Bennett <rossbennett34 at gmail.com>
wrote:

  
    
#
For this type of problem, I would probably generate one set of random 
portfolios and just reuse that set of feasible portfolios...  My usual 
rule is n-assets + 1-2k feasible portfolios.  You can get a better 
number e.g. from sampling theory, but this should be enough.

Once you have this weights matrix rp, you only need to regenerate rp if 
your universe changes.

Still interested in a more efficient implementation, of course, or we 
can work with you to see if we can find resources to work on it, e.g. 
from academia.

Regards,

Brian
On 03/20/2017 05:28 PM, Kevin Dhingra wrote:

  
    
#
Brian,

Yes I think that will be a good starting point. My universe would not
change a lot (I will be working with 10-15 benchmarks at a time and I guess
I can generate a reusable set for each independently before running it
through my main algorithm). Having said that, I envision the investment
mandates/constraints changing quite a lot (both in the cross section and
also over time for the same manager). I am hoping there must be a way
around it using rejection sampling but have not done enough research to
comment on how that solution works for such big dimensions. It will be
really helpful if you could point me to any specific resources from
academia for the same (Haven't been able to find much about random
portfolios myself except Portfolio Analytics and Patrick Burns work on
Portfolio Probe). As a side note - Do you think translating it using Rcpp
would be time well spent or you think there must be a smarter way to get
around it still using R?

I really appreciate your help on this thread.

Regards,
Kshitij Dhingra

On Mon, Mar 20, 2017 at 7:21 PM, Brian G. Peterson <brian at braverock.com>
wrote:

  
    
#
The process you describe is pretty standard for an investment-committee 
driven process.

I'm going to suggest that you don't really want to change the 
constraints that often.  For example, box constraints should be as large 
as your overall investment mandate allows to give you the greatest 
possible room for allocations.  Sector or Factor constraints likewise 
should be as minimal as possible just to guarantee the degree of 
diversification described in your investment mandate.

The reason I'm suggesting this minimal constraint set is one of the 
reasons we wrote the random portfolio code in the first place.  To see 
what I mean, generate a set of unconstrained random portfolios (or e.g. 
only with a full-investment constraint).  Then generate sets of 
constrained random portfolios, adding your various constraint sets. 
Plot the different sets on the same risk/return scatter plot, using 
different colors for each set.  Note how small the feasible space 
becomes, very quickly.

This shrinkage of the feasible space has some good shrinkage 
properties...  moderate shrinkage actually decreases the possible impact 
of estimation error in the various inputs, a little.  Large amounts of 
shrinkage (overly restrictive constraints) will do the opposite, and 
magnify the negative out of sample impact of estimation error.

The academic literature mostly focuses on analytical solvers (e.g. 
quadratic, linear, etc) and simple constraint sets.  We've cited papers 
by Patrick Burns as well as papers on the simplex models in 
PortfolioAnalytics, but the literature is not vast.

Numerical solvers become important as the feasible space becomes 
non-smooth.  One of the things that can create a non-smooth feasible 
space is a complex, overlapping constraint set.

The rportfolios package proposed by Frederick Novomestky also seems to 
be an R-only implementation, at a glance relying on truncated random 
binomial vectors rather than truncated random uniform vectors.  I 
believe it will have similar performance characteristics to the 
Burns-style random sample portfolios, and it seems to support fewer 
constraint sets (no overlapping sector, group, or factor constraints 
that I see).  In any case, it generates matrices of weights that are 
likely compatible with the PortfolioAnalytics random or seed portfolio 
inputs.  So if it works for you, that's great.

You also discuss using rejection after generating the portfolios.  This 
is the method used internally by random.portfolios to reject individual 
weights if a constraint is violated.  I'll have to evaluate whether the 
truncdist package used by rportfolios could be more efficient than the 
runif that is used by the current code.  PortfolioAnalytics also allows 
portfolios to be penalized in the solver, so that more complex cases can 
be considered, or interactions between constraints and objectives.

To answer the question of whether Rcpp will help is somewhat complex. 
I'm confident that some of the nested loops in the generation code will 
be sped up by Rcpp.  It is possible that more efficient algorithms are 
available for constructing the weight vectors.  A reason that this 
hasn't been a huge priority though is that construction of the random 
portfolio matrix is usually not the time limiter in a large 
optimization: your objective function is.  I think it will be possible 
to improve the efficiency of this step, though it is unclear how much of 
an impact this should have in practice to a large and complicated 
numerically solved portfolio optimization problem.

Regards,

Brian
On 03/20/2017 07:06 PM, Kevin Dhingra wrote:

  
    
#
Just wanted to point out an example of what Brian mentioned for
visualizing the feasible space with different constraint sets.
Visualizations here
http://rossb34.github.io/PortfolioAnalyticsPresentation2016/#22 on
slides 22-26 and code
https://github.com/rossb34/PortfolioAnalyticsPresentation2016/blob/master/feasible_space.R
to produce the plots.

Ross
On Tue, Mar 21, 2017 at 5:41 AM, Brian G. Peterson <brian at braverock.com> wrote:
#
Brian, Ross and Frednovo,

This has been extremely helpful. I think now I have a lot better
understanding of how to think about the problem at hand. Appreciate it.

Regards,
Kshitij Dhingra

On Tue, Mar 21, 2017 at 7:24 AM, Ross Bennett <rossbennett34 at gmail.com>
wrote:

  
    
#
On Tue, Mar 21, 2017 at 5:41 AM, Brian G. Peterson <brian at braverock.com> wrote:
<snip>
Brian has hinted at this, but I want to say it explicitly.   Whether
or not moving to compiled code is worth it is mostly an empirical
question.  And it's difficult to do more that speculate, unless you
have profiling data.  So I would strongly encourage you to profile
your optimization before you change any code.  I would be happy to
help review the profiling output.
<snip>

  
    
3 days later
#
FYI- I have found it very useful to use random portfolios to verify that
your optimization is actually doing something. Particular mandates get so
bogged down by constraints (liquidity, box, sector, industry, Barbra risk,
etc) that every answer in the feasible set is basically the same.  (i.e. If
you are making money it's on the constraints.)

- Scott
On Tue, Mar 21, 2017 at 13:43 Joshua Ulrich <josh.m.ulrich at gmail.com> wrote:

            
1 day later
#
I'm coming significantly late to the thread, but I'll throw in a bit of 
my view on the required number of random portfolios to generate.

I think it depends a whole lot on what you are doing.

If you are essentially looking for a p-value, then you'll need a few 
hundred at most.  You shouldn't really care if the p-value is 4.3% 
versus 4.1%, and if you're looking at tiny p-values, then algorithmic 
limitations and the subtleties of what "random" means in the particular 
situation are going to start to have big impacts.

If you are looking at densities, then something like 10,000 portfolios 
is not terrible.  If you are looking at tails, then you want lots more. 
But again, you have lots of model risk in the tails.

Pat
On 20/03/2017 23:21, Brian G. Peterson wrote: