The essence of optimisation

New comment
 

For which types of strategies does optimisation make sense?

 
toxic:

For which types of strategies does optimisation make sense?

For all those that have numerical parameters.
 
micle:
For all those who have numerical parameters.
the truth is spoken by the man. in short and to the point. parameters are an infantile bunch of lazy numbers. they lack discipline and order. so they are sent to the optimization cauldron to learn where they belong.
 
micle:
For all that have numerical parameters.
nowi:
the truth speaks man. briefly and clearly. parameters are an infantile bunch of lazy numbers. they lack discipline and order. that's why they are sent to the cauldron of optimization to learn where they belong

Hmmm, I used to think so too, but lately I've been having doubts and the "truth" has been slipping away.

I don't want to produce too many "letters", I'll try an example.

For example: Random entry, TP and SL.

Let's optimize TP and SL.

For simplicity, we may fix the proportion ofTP / SL and optimize one parameter (their multiplier).

Does the result of such optimization make any logical or practical sense, and the main thing is WHY?

What does it depend on?

 
toxic:

Hmmm, I used to think so too, but lately I've been having doubts and the "truth" has been slipping away.

I don't want to produce too many "letters", I'll try an example.

For example: Random entry, TP and SL.

Let's optimize TP and SL.

For simplicity, we may fix the proportion ofTP / SL and optimize one parameter (their multiplier).

Does the result of such optimization make any logical or practical sense, and the main thing is WHY?

What does it depend on?

This optimisation makes no logical or practical sense simply because the input is random....
 
toxic:

For which types of strategies does optimisation make sense?

Not for all strategies, but for most scalpers optimization makes no sense, because the results in the tester and on the demo (real) differ very significantly.
Moreover, scalper results in the tester depend on testing modes: normal mode, with random delay, all ticks, OHLC on M1.
 
Trend-following strategies tend to give fairly reliable results. Therefore it makes sense to test and optimise such strategies.
 
micle:
This optimization makes neither logical nor practical sense only because the input is random....

That's what I mean. It turns out that optimization does not make sense for all strategies with numerical parameters.

Let us continue our reasoning.

If not for all, then for what strategies it makes sense and for what it does not? Can we not only list all possible variants of strategies or their low-level technical characteristics, but also generalize them into a reasonable number of categories?

So the first stated property leveling the meaning of optimization is randomness of input. Why is it so? Technically, on the MO of each transaction profit, entry and exit affect equally, i.e. at random entry and quality exit we have approximately twofold decrease of MO of transaction profit in comparison to quality entry and exit. But the constant TP\SL as it is clear does not pretend to be a "qualitative" output, though it is possible to select a "nice" one with acceptable retournaments and other statistics by optimizing even in this way from a set of tests.

What is wrong with it? Let's put a more general question: How does the quality and dryness of optimization depend on the "quality" of inputs and outputs? How can this be quantitatively interpreted?

Let's take the most popular "МАшки" strategy as an example

Strategy: The signal to open is a crossover of the fast moving average over the slow one, from bottom to top, and to close from top to bottom. Always in the market, roll over.

Question: optimizing the moving window, separately or together proportionally, may give fundamentally different results than in the first case (random+ TP\SL) ?

pagot:
Not for all, but for most scalpers optimization makes no sense, because the results in the tester and on the demo (real) are very, very different.
Besides the results in the tester depend on test modes: normal mode, with random delay, all ticks, OHLC on M1.

Let's leave the topic of velocity, especially in the context of the Metatrader's tester, but not because of fundamental differences in testing and optimization, but because of the absence of a real tape (ticks) or data less than 1 m only because of it. In essence the difference is the same as inability to test intraday strategies on dailies, if the dailies were the minimum step on which we may fix.

In the context of exchanges, there are also questions of liquidity and real slippage, i.e. not lag or DC-synthetic, but real eating of the stack.

pagot:
Trending strategies tend to give fairly reliable results. So it makes sense to test and optimize such strategies.

We are talking exactly about optimization, the testing itself is a trivial process. It is about validity of extrema in the parametric space and what their validity reliability depends on.

 
Wow, a potentially interesting topic for once.
 

toxic:

...WHY?

What does it depend on, the presence of such meaning?

I have a number of thoughts about this, if I understand correctly, but they are so vacuum and impertinent that it would be a spit in the eyes of the whole TA community, so I will keep the details to myself.

If in general we can say that effectiveness of optimization directly (but non-linearly) depends on effectiveness of predictive model, the more "random" model the more optimization is a fitting.

A perfect model doesn't need optimization or it is part of model and has no external parameters except for some risky coefficients (aggressiveness), though unlikely, there can be strictly black box, absolutely black.

In my opinion it is the same as optimizing the noise strategies by optimizing noise .

So everything is simple, you check the prediction ability of the model and simultaneously get both its absolute meaning for any further manipulations and its optimization potential.


P.S

I can predict your further iterations of complexity of structuring strategies asking how they differ in terms of optimization and in advance I can say that even neural networks applied to for example price or increments, fit as pure random with parameter in the form of seed.

 
toxic:

For which types of strategies does optimisation make sense?

The question is not quite accurate.

Before you can develop strategies, you need the right idea that gives a positive result. There are many strategies that can be developed based on this idea, with sufficient difference in terms of the design. It is necessary to evaluate and select the best strategies in terms of profitability and reliability. And the final step, after all this, is to optimise the parameters of these strategies.

There is no point in optimising a strategy that has the wrong idea. This may result in a particular version of fitting to the story, with no prospects for the future.

123456789101112
New comment
Reason: