Criterion for automatic selection of optimisation results. - page 3
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
ivandurak писал(а) >>
А как же распределение результатов сделок . Львиная доля прибыли может быть и в начале исследуемого периода . Имхо для начала необходимо договорится о критерии по которому будем отбирать варианты оптимизации, а иначе флуд и опять плевки на спину .Лично мне импонирует - близость результатов торговли к прямой линии на постоянном лоте, но не настаиваю .
I agree with the highlighted blue.
With red highlighted, no. A straight balance line is not an indication of stable results, especially if traded with a fixed lot. A straight balance line can also be obtained by a variable lot size dependent SL, expressed as a percentage of an agreed upon initial capital amount (not as a percentage of the current balance)
I will try to give some thoughts on the topic of the branch a bit later.
I don't think there is enough to do with one 'criterion', there needs to be a balance of criteria.
Maybe so. Probably a set of criteria is better than one. But from the practical point of view, it would be desirable to have one criterion that is as capacious as possible. There are no obstacles to analyse it in combination with something else, aren't there?
Here is the simplest variant that works well (much better than standard variants, in terms of speed and results) in my optimizer:
First, filter by number of deals, then simply maximum GP * P * PD/(GL+MD); // it was derived from nothing, just intuitively, after that I started thinking about a branch
-Profit = P,
-GrossProfit = GP ,
-GrossLoss = GL,
-MaxDrawdown (Drawdown) = MD ,
-Number of profitable trades = PD, losing trades = LD, Total number of trades = AD,
-Number of bars (ticks) of testing = TIME,
-Max Profit trade = MPD, max loss trade = MLD
-series of profitable trades = SPD (in units), = SPD$ (in deposit currency), series of losing trades =SLD, =SLD$
Profit, MO, profit factor. relative drawdown - these characteristics are calculated at the end of optimization, when the position is forcibly closed.
Maximum drawdown - a characteristic that we can use as a risk extremum. Optimization can end when the balance line is in the horizontal channel formed by the maximum profit (unfortunately, this characteristic is missing in the report) and the maximum profit minus the maximum drawdown. I have adaptation mechanism of the Expert Advisor above.
The purpose of optimization is to calculate the risks. Knowledge of possible risks allows us to apply MM.
1. Out Of Sample-- i.e. data (quotes) that the TS has not seen and the results of the work on that data;
2. Expectation - average profit from a trade in pips or selected currency;
3. Time series (read - quotes).
Cotier can be decomposed into its constituent waves, which have different frequencies and amplitudes (I say this for clarity, not hinting at a practical decomposition using Fourier transforms etc). This can be seen with the naked eye (waves, Eliot has nothing to do with it). Every TS, written or to be written by anyone, is an attempt to describe the behavior of a specific individual wave. And since the frequency of waves is different, the number of deals in different TS will be different, may be a lot, may be little.
It seems to me that one "criterion" is not enough, we need a certain balance of criteria.
And it seems to me there is not only one "criterion" but some hypothetical balance of criteria as well.
I will draw an analogy with the animal world (excuse me, but my brain operates with "biological" images more easily) :)
Each constituent kotir wave is the natural habitat of a certain species. Some live in a flat and sparsely vegetated desert and others in a dense jungle. These species have completely different ways of making a living and the rational use of the energy they extract.
Let's go back to the TC again. How can we determine which ones are worthy of working with our blood and which are not? Based on one, albeit universal, criterion?
I am increasingly inclined towards the idea of using multi-criteria optimization, each criterion is a description of a certain population, in terms of GA. Thus, several different populations of different species of individuals will fully exist, giving an opportunity to interbreed with individuals from different populations, it is possible to strengthen the best qualities of representatives of different species by means of GA.
PS It remains to describe the individual TC species in terms suggested by FigarO. What is the most difficult for me, I am not familiar with mathematics. I have made a similar request to one of the forum members, but either I could not interest him, or I made it at the wrong time.
PPS The formulation of the fitness function is almost more important than finding an input dataset for the NS, this also applies to any TS.
I will add my opinion on the question.
1) There is no way to do without set of parameters, these parameters are taken from the tester, you can add some, but unnecessary ones will also need to be eliminated, as unnecessary inputs in neural networks(analogy).
2) Outputting general coefficient by means of multiplication most likely will appear new sets of local minima, or these criteria will essentially drive GA to false (not necessary for us) minimum/maximum (there is no difference). So I recommend just forget about the multiplication made by the author of the branch a few posts above. I experimented with them, it does not work out well.
3) Essentially, the task is "multi-criteria optimization" or optimization according to many criteria simultaneously. All I've read on this subject: everything boils down to giving each criterion some weight[0;1] and using a simple linear equation I can deduce this general criterion (Y = a1*x1 + a2*x2+...an*xn) - this will also lead to false minima in our case, or we will have to conduct a separate investigation (for every strategy) of the surface first, look how weights influence minima/maxima, whether GA goes to a false extremum (false from our point of view) - generally speaking, this is not the way either.
That is why you (I appeal to those who are interested in the branch) will have to invent a new method that will not be in the books.
Problem: the first thing you need is to get rid of dimensions, for example percentage of profitable transactions in the range [0;100], while your profit is theoretically unlimited, etc. Let me tell you right away that scaling will not work (it is the same as assigning weights to indicators), there will be many outliers that will dump GA into a false minimum.
Можно критерий оптимизации получить методами генетического программирования, но боюсь, пока вычислительных ресурсов писишек не хватет, хотя разрабы обещают облачные вычисления...
Everything brilliant is simple... so I don't think it's necessary to set superlatives...
Everything brilliant is simple... so I don't think there's any need to set superlatives...
I agree. Complicating things doesn't guarantee better results.