The essence of optimisation - page 2

New comment
 
papaklass:

Optimisation answers the question: Whichvalues (numerical) of the parameters have the best values at the studied site?

We did the optimization, got the values, so what? These "best" values of the parameters in an area other than the investigated one will give different results.... :)

I'm not optimizing at all, as I find it meaningless.

I check my strategies for trading appropriateness in one way - I change position opening direction to the opposite one (for example Buy to Sell). If the MO of the strategy remains positive, I implement it.

+1

As for me, I think that even the MO can be rid of as a rudiment, or rather as a factor having quite a remote effect on the trading results, with targeted weighted averaging. We are dealing with Wiener process(W), "martingale" and therefore we can control only risks conditionally.

My prescription - GOOD SATISFACTION, soft martin, and you'll be happy. And the "forecasting" is left to all kinds of hippies. Only sick in the head or crooks think you can "predict".

 
Has anyone tried to correlate the robustness of the optimised parameters and the results of the optimisation with numbers?
 
TheXpert:
Has anyone tried to correlate the robustness of the optimised parameters and the results of the optimisation with numbers?
What is there to try? Strategy optimization is essentially a statistical estimation problem. So you can use the well-known methods to calculate confidence intervals for the parameters and then check the performance of the system in the intervals obtained.
 
anonymous:
So it is possible to calculate confidence intervals for the parameters using known methods, and then check the performance of the system in the resulting intervals.
How? Given that optimization and estimation are performed on different data and strategy statistics on the forward and backtest do not have to coincide?
 
TheXpert:
In what way?

http://quantile.ru/03/03-SA.pdf

Given that the optimisation and evaluation are done on different data and

Obviously - estimate parameters on one part of the sample and test on the other.

The strategy statistics on the forward and backtest don't have to match?

Limit yourself to considering only those systems that work out of sample and don't waste time on the rest.

 
anonymous:

Limit yourself to considering only those systems that work out of sample and don't waste time on the rest.

No, you probably don't understand. A system on OOS can work but have strikingly different results from the backtest.
 
TheXpert:
A system on OOS can work but have strikingly different results from the backtest.
I understand this very well. My statements do not contradict this.
 
anonymous:
I understood it perfectly well. My statements do not contradict that.
So you are suggesting that such systems should not be considered immediately, even if they work?
 
TheXpert:
So you propose not to consider such systems right away?

There is also no point in abandoning such strategies altogether, as they can serve as a basis for another strategy, or improve the ones already in use. But they are of little interest for trading.

 
toxic:

It turns out that not all strategies with numerical parameters can be optimised in a meaningful way.

Whether it makes a difference or not, if there is a set of numerical parameters, then you still have to choose one or the other.

for example for your strategy you need a short waving machine. which one do you take? 2,3,4,5 ? And in what range can it still be considered short?

Another thing is that if the trading idea is good, then it should not have a too large range of numerical parameters to choose, otherwise this rather shows the absence of the idea or its incompleteness. for example if you know that you need a short Mach. you won't check the period of 200 and will be limited to a range (2.....10). but you still have to choose something specific
123456789...12
New comment