strategy optimisation

[djames]

for giggles i'm using sklearn and a genetic algo. library to optimize parameters for my trading strategy.

in order to rank the sets of parameters i need a single valued result from the optimization, e.g. Sharpe ratio, but of course there are many other statistics which are not represented by Sharpe ratio.
How do you go about combining these statistics into a single value?

 

[Simples]

for giggles i'm using sklearn and a genetic algo. library to optimize parameters for my trading strategy.

in order to rank the sets of parameters i need a single valued result from the optimization, e.g. Sharpe ratio, but of course there are many other statistics which are not represented by Sharpe ratio.
How do you go about combining these statistics into a single value?

What exactly do you want to get out of your trading?

Many ways of combining values and often naive approaches may work well, but it helps to sort out what kind of results you really want first, before designing a fitness function.

 

[djames]

just wondering if there are standard combinations of return statistics to give the fitness output value

 

[Simples]

just wondering if there are standard combinations of return statistics to give the fitness output value

That'll depend on your data and what you want to do with them. If you always have the same data, you might want to multiply the values. If you need, you may want to normalize them first and/or make sure they're the same sign. The more you do with the values, the more information you lose though. If you want some optimal selection/combination, you may try ranking and/or normalizing that. If you think there's some optimal standard, then you close all the other possibilities.

 

[wintergasp]

for giggles i'm using sklearn and a genetic algo. library to optimize parameters for my trading strategy.

in order to rank the sets of parameters i need a single valued result from the optimization, e.g. Sharpe ratio, but of course there are many other statistics which are not represented by Sharpe ratio.
How do you go about combining these statistics into a single value?

I'm not sure I understand @Simples response.

The sorting factor for your optimization results is THE main secret sauce of any machine learning system, it's what allows you to avoid over-fitting and have an adaptive behavior by being able to automatically changing algos.

So..... won't tell you want we use but I can tell you you're on the right path to a robust system, a lot of guys are far from there.

 

[Simples]

I'm not sure I understand @Simples response.

The sorting factor for your optimization results is THE main secret sauce of any machine learning system, it's what allows you to avoid over-fitting and have an adaptive behavior by being able to automatically changing algos.

So..... won't tell you want we use but I can tell you you're on the right path to a robust system, a lot of guys are far from there.

I mostly replied to this question: How do you go about combining these statistics into a single value?

One value to rule them all, that'll depend on what one want to get out of the trading: Ie. how capitalized you are, risk tolerance, can you live with larger drawdowns if it means even larger wins, how long can you stand to wait, etc.

Not sure there's a perfect answer. The design will depend what one is looking for in a trading system.

 

[dtrader98]

I'm not sure I understand @Simples response.

The sorting factor for your optimization results is THE main secret sauce of any machine learning system, it's what allows you to avoid over-fitting and have an adaptive behavior by being able to automatically changing algos.

So..... won't tell you want we use but I can tell you you're on the right path to a robust system, a lot of guys are far from there.

@wintergasp
I like that response. Could you elaborate a little more?

@djames
There are ways to combine by importance weighting, but it might be better to start by clearly understanding your objective and constraints and building an objective function from there (like simples pointed out). Doing it that way helps you to understand shortcomings and how you want to modify them.

I recall there's a blog called QUSMA with an example of combining objectives in a sabermetrics kind of way (TRASYCODRAVOPFACOM). Ok, here's another blog combining objectives into a bliss function. Both are examples of combining standard trading objective functions-- that might be what you are looking for.

 

[931]

The sorting factor for your optimization results is THE main secret sauce of any machine learning system, it's what allows you to avoid over-fitting and have an adaptive behavior by being able to automatically changing algos.

I think filtering sauce is interconnected to algo/ parameters and end result is all sauces mixed. Even with many parameters and powerful hardware need lucky or smart solutions and more possible combinations can also introduce more overfitting problems that are hard to remove or limit by overcomplicating code further.

It probably all comes down to luck on ideas and time ,effort u are willing to put in testing dif ways.