Combining multiple systems

Quote from phattails:

Just to reiterate what TSGann said and to further organize.
Goal: define a methodology for combining multiple systems

Objectives:
a. identify what separates "a system" from multiple systems.
1. differences... commonalities
2. types of simple (if there is such a thing)
b. define ways to combine systems
1. define end goal for combination
2. quantify (they don't have to be old)
a) define assumptions of correlation
b) reinvent correlation or
c) sanity check: can we have a measure that's reliable?
i. some of the time (regime switching)
ii.
d). lit. review on ltcm, sub-prime, etc.
3...

Keep in mind, as I alluded to in response to one my first interviews ever "to know where I will be in ten years is to assume all of my oppurtunities, future values and joys are fixed and known... if I made this assumption, just think think how I could stifle this company's philosophy of innovation". In other words our goals and objectives are subject to change.
More...

:confused:

You're over-analyzing my original questions. It has little to do with the actual process... It's simply a question to be discussed.

So the Objective List you made does not reflect my point of view. Though, I think it's a good starting point for a discuession.

:D :D :D
 
Quote from TSGannGalt:

:confused:

You're over-analyzing my original questions. It has little to do with the actual process... It's simply a question to be discussed.

So the Objective List you made does not reflect my point of view. Though, I think it's a good starting point for a discuession.

:D :D :D
More...

I didn't use your post as a starting point, but the first objective I came up with happened to be your question. I quickly assumed that my direction was your direction. I guess I was simply being curtious and structured it the wrong way.

sorry
 
Well... things have been very conclusive (in a sense) in this thread...

In terms of, developing single signals and position sizing, I think there's been plenty of discussions about it. But there hasn't been enough discussions about portfolio management from a systematic trading standpoint. I've done some research regarding it and continue to do so, and I'm sure there are others (though very few in ET).

So...

I'll start again with a question... (again)

What risk measure"S" should traders be using to measure they're model? From a single system to a portfolio standpoint?

To explain a bit with my newbie language:

Sharpe ratio is obviously not the perfect measure to use as a fitness. It's not almighty. A simple example would be when you develop a good trend-following model. Using Sharpe would thrown it out easily. And obviously, you can have a piece of shit intraday system but it can survive the selection criteria.

Well....

--------------------------------------------------------
To narrow it down a bit more...

Starting with a single system, what type of measures will be useful to measure a single system, from a portfolio standpoint?
- Would you, also, throw out a good, yet volatile system, or implement a crappy but low volatility system and why?

Obviously, the simple approach would be create a profile of your current portfolio, then run a diff. analysis, but hopefully we can keep that out for my question above.

PS. Alan, I know you read this. I'd love to hear from you about this topic.
 
Quote from TSGannGalt:

I'll start again with a question... (again)

What risk measure"S" should traders be using to measure they're model? From a single system to a portfolio standpoint?
More...

Hey, why do you ask questions? Could you share your own results?

I tried to find "optimal fitness function" for my strategy. Not portfolio, but single system. I have tried Sharpe, yield/max DD, Sortino, Lottery approach (stochastic dominance I and II).

But I do not differ the best fitness function. Sharpe is not good for me. May be simple profit factor is quite good as an optimizational parameter. Or I mix together some formula

fitness = PF * (yield / max dd) ^ 0.1
so on
From mathematical point, the solution is the Stochastic Dominance II.


Any aproach leads to overfitting if you do not have enough data.



Quote from TSGannGalt:
Starting with a single system, what type of measures will be useful to measure a single system, from a portfolio standpoint?
- Would you, also, throw out a good, yet volatile system, or implement a crappy but low volatility system and why?
More...

I don't think that volatile system will have a good Sharpe ratio.
 
There are a lot of "measures of goodness" (also called "fitness functions", "Gain-to-Pain ratios", "performance statistics", etc).

Ask yourself "why?"

Why are there dozens and dozens of different ways to evaluate the desirability of an equity curve? With all of these to choose from, why do people continue to invent new ones, even today?
  • Sharpe Ratio
  • Sortino Ratio
  • Ulcer Index
  • Kestner K-Ratio
  • MAR Ratio
  • Return Retracement Ratio
  • Semideviation Sharpe Ratio
  • R-Squared
  • Sterling Ratio
  • Profit Factor
  • Seykota Lake Ratio
  • Pessimistic Return On Margin
  • Correlation to Perfect Profit
  • Walk-Forward Efficiency
  • R-Squared
  • Monte Carlo derived, 95% confidence, CAGR
  • WL Score
I believe the answer is: because no two traders can agree on what constitutes a "desirable" system. So each trader invents his own measure of desirability .... which all other traders reject.
 
Quote from MAESTRO:

Read about Parrondo's paradox in the link I provided ^^ above. There are some examples of the games that are individually losing but in combination they produce positive expectations. It took me a while to wrap my mind around it. I actually have built a system based on this paradox and tested it on ES markets. I got very surprising results! Totally counterintuitive.
More...
you hooked me in. i can hardly believe this ...
 
overall on the topic of this thread.

little or even better negative stress correlation is what
you want. who cares about correlation when things go
up? that is the whole problem about markowitz. he looks
at simple correlation, which, in upward sloping trading
systems will most of the time be: corr in upward e-curves.
which is the one you don't care about.

i dare say the real problem does not lie in the computation
of correlation but somewhere else: how to develop
another strategy which is really different.

IMO this takes a lot of discipline. you really have to dig
at another corner. maybe even take other markets.
or other time frames. it is so difficult not to just trade on
the same effect, just by means of other tools.

the problem is that system development as such is
real hard work and it is very tempting to lean back,
once you finally have found something that works.
it is difficult to motivate yourself to do it all over again
somewhere else.

so the real problem takes place in the mind.
 
Quote from TSGannGalt:... Sharpe ratio is obviously not the perfect measure to use as a fitness. It's not almighty. A simple example would be when you develop a good trend-following model. Using Sharpe would thrown it out easily. ...
More...
our take: we discuss what next development would
increase overall sharpe highest per unit time of development.
we trade futures at the moment, both on daily and
intraday data. we decided to look at stocks, using
variations of existing methodology. hopefully uncorr
and not too much time of development.

the problem is that you need to judge this before you
develop. we finally looked at wavelets six months ago.
took us quite some time to find out that they by
themselves do not make much sense. that is the
problem with the complex stuff: you really pay
development time for this additional potential sharpe.

this illustrates my earlier point: additional sharpe per
unit of development time. it could well be that you have
a set of strategies and long term trend following, while
having a spare of 0.8 on its own is the best thing to
develop, since it might be uncorrelated to your stuff
that is on trading and it does not take too long to
develop.

i honestly don't believe too much in overquantifying
the issue. with every backtest, which is essentially
the basis of your correlationMonteCarloVaRWhatHaveYou
analysis you have the danger of overfit. that is the
biggest issue in my eyes. and if you put optimisation
on top of that ... well, most of the time just waste
of that very time.

all IMHO. and i am not jim simmons. ... :)
 
Quote from man:

you hooked me in. i can hardly believe this ...
More...

The more you learn about it the better it becomes. Don't listen those "know-it-all" guys here. I guarantee that if you investigate it thoroughly you will find it more than rewarding.
I stopped answering to stupid remarks long time ago. I am on the side lines these days. Just enjoying the science.

MAESTRO
 
Quote from MAESTRO:

The more you learn about it the better it becomes. Don't listen those "know-it-all" guys here. I guarantee that if you investigate it thoroughly you will find it more than rewarding.
I stopped answering to stupid remarks long time ago. I am on the side lines these days. Just enjoying the science.

MAESTRO
More...
i stopped believing in anything i hear and read until i
tried it out within my team ... :) ... but usually my guys
come back to me: "all crap" ... but who knows, once in
a while ...

my take on this, just from top of my head and without
looking at it: the systems, even if they do loose over
years must have certain features. they can't be random,
since out of random you can't moneymanage return.
martingale blablabla.

this certain feature must be autocorrelation. (sorry
you probably covered all that ... i did not bother reading
the whole thread ... too weird a day out there ...).
now your system of systems just trades that: the
autocorr of the underlying equity curves. all of a sudden
it makes perfect sense.
 
You must log in or register to reply here.
Back
Top