Optimal portfolio of trading systems

daskapital · Post by **daskapital** » Fri Jun 19, 2015 12:40 pm

Suppose that I have N trading algorithms generating N asynchronous return series. How do I work out the optimal mix relative to, say, Sharpe ratio?

I could, of course, interpolate all data series to a uniform time step (the greatest common divisor), and go with the standard portfolio optimization algorithms. But wouldn't this naive approach tend to favor the system with the longest holding periods? A positive return on, say, daily, interpolated on the hourly, would produce a nice smooth uptrend, which of course seldom is true.

Ideas?

deeds · Post by **deeds** » Fri Jun 19, 2015 12:50 pm

The framework in this paper may be relevant, setting aside the immediate application described by Derman.

mmport80 · Post by **mmport80** » Fri Jun 19, 2015 1:01 pm

Equalise marginal risk contribution of each strategy? Best bang for your risk adjusted buck..

daskapital · Post by **daskapital** » Fri Jun 19, 2015 1:50 pm

Thanks for the suggestions. I'll look into Derman's paper, deeds.

Mmport - any suggestions for practical implementation? The returns are asynchronous, and estimating correlations in such signals is not trivial (e.g. http://arxiv.org/abs/0805.2310).

Rashomon · Post by **Rashomon** » Fri Jun 19, 2015 3:34 pm

daskapital, I'm not sure I understand you correctly but it sounds like a multi-armed bandit problem.

mmport80 · Post by **mmport80** » Sat Jun 20, 2015 4:00 am

Who are your stakeholders? What's their time frame?

I doubt anyone care about their hourly return rate, right? Perhaps they focus on daily or longer?

If that's the case, perhaps the async-ness is not the issue you believe.

Then again you know your situation better than me, paint a picture, perhaps I'll understand better then...

Scotty · Post by **Scotty** » Wed Sep 16, 2015 5:28 pm

Hi DasKapital

Following mmport80's line of thought, see the book 'Introduction to Risk Parity and Budgeting' by Roncalli.

Also see the 'alpha streams' stuff from Zura Kakushadze (via SSRN).

Let me know how you go or if you want to discuss more.

Scotty

NeroTulip · Post by **NeroTulip** » Thu Sep 17, 2015 2:40 am

Interpolating returns to the highest frequency would give garbage because of the smoothing effect you mention. On the other hand, you would not go terribly wrong aggregating to the lowest frequency. If you have hourly and daily strategies, the correlation of daily returns would not be a bad starting point.