Criterion for automatic selection of optimisation results. - page 2
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
We struggle, we write strategies, and virtually any expert is capable of generating profits with certain parameters on the final trading interval. Meanwhile, relatively little attention is paid to the selection of parameters.
I don't seem to have missed anything (correct me if I have), all kinds of MO, PF etc. are omitted deliberately, as they are calculated and can be obtained from the above.
Well, let's do some magic.
We have somehow forgotten about a wonderful green image available in the optimizer. If it's all green and about the same shade, then it's a nice CU. It can be a fairly convincing basis for this.
Adapting the advisor to the current market conditions takes place in 3 stages:
1. optimisation
2. Forward testing to the current moment
3. Correct connection of the Expert Advisor.
We would like to discuss the 1st stage. What do we expect from it? Finding the parameters of the Expert Advisor that give higher profitability and lower drawdown. The recovery factor takes into account both of them. But there is something wrong anyway. There is no stability of the result. That is the main problem.
What is characteristic of stability of an Expert Advisor? First, it is limited risks relative to the initial balance or, in other words, the value of the maximum drawdown is finite. Second, it is a small series of losing trades.
Adapting the advisor to the current market conditions takes place in 3 stages:
1. optimisation
2. Forward testing to the current moment
3. Correct connection of the Expert Advisor.
We would like to discuss the 1st stage. What do we expect from it? Finding the parameters of the Expert Advisor that give higher profitability and lower drawdown. The recovery factor takes into account both of them. But there is something wrong anyway. There is no stability of the result. That is the main problem.
What is characteristic of stability of an Expert Advisor? First, it is limited risks relative to the initial balance or, in other words, the value of the maximum drawdown is finite. Second, it has a small series of losing trades.
BP is non-stationary and we cannot speak about stability. It's better to speak about stability of TS according to the updated market conditions. It seems to me it is equivalent to the stability of testing results in relation to changes of TS parameters. This is what the results of testing in the three-dimensional image show (Elena picture). If, roughly speaking, you have exactly one green square. then you have found a pimple at the bottom of the pool and nothing will help your TS. But if the whole chart is green and the intensity of colour changes slightly - you have a stable TS and changes in unsteady BP will not lead to a draining of the deposit.
At the first stage we find variants that satisfy the condition of maximum drawdown less than or equal to N% of the initial deposit or loss series less than or equal to K trades.
At the second stage we check stability of the variants found by prolonging the time interval of optimization up to the current moment.
At the third step: connect the Expert Advisor correctly. What does it mean?
Let's review the structure of the Expert Advisor:
1. Signal to buy/sell.
2. position opening.
3. A signal to close a position.
4. Close position
5. Analysis of transactions history.
When we test, this chain is followed automatically. In order not to artificially break the market conditions, we should provide an order of actions when connecting, i.e., the Expert Advisor starts at point 1 after point 5 is completed...
Шарпы, Сортино и прочие - уже почти птичий язык:) Все говорят круто, но я тоже не пробовал, если кто знает как это рассчитать из исходных данных для практического применения, попробую и скажу как результат. Надо переобозначить исходные данные для формульного языка.
-GrossProfit = GP,
-GrossLoss = GL,
-MaxDrawdown (Просадка) = MD,
-Колличество прибыльных сделок = PD, убыточных сделок = LD, Общее количество сделок = AD,
-Колво баров(тиков) тестирования = TIME,
-Макс. прибыльная сделка = MPD, макс. убыточная сделка = MLD
-Серия прибыльных сделок= SPD (в штуках), = SPD$ (в валюте депозита), серия убыточных =SLD, =SLD$
Ничего я не упустил? Можно рисовать формулы?
З.Ы. Возьму пару часов для раздумий, и морально поддержу наших олимпийцев, покатаюсь на лыжах) А то без моей поддержки) они пока не очень....(
Кто со мной?:)
What about the distribution of transaction results . The lion's share of profits may be at the beginning of the period under study. Imho first we need to agree on the criteria by which we will select the optimization options, otherwise the flood and again spit on the back. I personally like the proximity of the trade results to the straight line on a constant lot, but I do not insist.Whether the descriptions and formulas will be useful for research I do not know, but I copied them here, maybe there will be ideas how to apply....
I think they will be useful, thank you, it's convenient to have everything at hand, but we need a good "interpreter" in human language, like Mathemat, to lure him here. Maybe he with his knowledge of mathematics could land these formulas for practical use.
I see that while I was doing my skiing exploits, the discussion took a bit of a wrong turn)
It's all true in principle, and BP is unsteady (I've been using synthetic lately, by the way), and it's possible to get a fit, and OOS is the head of everything, and TC stability is very good... All so, but how to use it practically in context of statement of question?
We have results of tests (optimization), there are a million of them, we have to choose several from analysis of list of figures on the first page. In principle, this list can be expanded by what can be calculated having a complete test result.
I see a few options, the first:
- Took a list of results, sorted by max profit - the worst half of the results to the furnace, then sorted by MO - the worst half of the results removed, and so on. (drawdown, profitability, number of deals...) If necessary, repeat the cycle of selection until the required number of results is left. For example, as far as I remember it was done in the first version of xeon' s automated optimizer. I think this approach has some disadvantages, and first of all they are connected with defects of parameters: MPO, Profit, Drawdown, etc. Of course, they contain some amount of information. Of course, they contain some information, but they are rather one-sided. For example:
MO is an average profitable trade, what can it give without taking into account the number of trades? One huge deal would be cooler than anything else...
Profit - without taking into account the profit factor? Who wants huge profit with low profit factor?
Drawdown - without profit factor? The drawdown is 0 and profit 1, do we need it?
And what happens when these parameters are set as a criterion for the GA of the terminal optimizer? I cannot even imagine. It's true, maybe there is some know-how there, but I don't know about it.
And that's why the second variant: to create a criterion that will take into account all necessary for us to the necessary extent, and the best variant according to this criterion with the probability higher than acceptable will meet the properties we are interested in (stability, profitability of GSO, etc.)
А потому, вариант второй: создать критерий который в нужной нам степени будет учитывать все необходимое, и лучший согласно этому критерию вариант с вероятностью выще приемлемой будет отвечать интересующим нам свойствам (устойчивость, прибыльность ООS и т.д.)
I don't think there's enough to do with one 'criterion', but there needs to be a balance of criteria.
What about the distribution of the results of the trades . The lion's share of profits may be at the beginning of the period under study . Personally I like the results of trading close to the straight line on a fixed lot, but I do not insist.
I agree, the factor is not insignificant. Can you show me how to practically calculate, put into mathematical form this distribution having a complete test result?