Criterion for automatic selection of optimisation results. - page 6

New comment
 
Figar0 >>:

Т.е. если я правильно понял, как бы оценивать не результат, а качество сделок, насколько сделки отвечают тому, что заложено в систему?

Надо покрутить в голове.

In general, yes. The main thing is to be able to see a picture of the trades as much as possible. Even if the filter misses dangerous trades, you should mark them somehow and see the chances of their triggering. If the filter moves the risks on the probability of probability of deal triggering closer than the main signal's radius, then this filter will set up in the future. And it's bad if it eats a large part of profitable ones. It's easier to open everything that is within the nearest opening radius of the signal and try to keep it. The main thing is to clearly define the radius.

 

If the TS is well profitable, all possible trades are opened and only the drawdown spoils the picture, then there is a cure for this disease too...

The main thing is to have all trades in front of your eyes...

 
Here's a reflection on the subject, by the way...
 

Read on, read on...


I forgot to tell you, the formula I gave is only applicable with a proportionally increasing lot.

 
ivandurak >>:
А как же распределение результатов сделок . Львиная доля прибыли может быть и в начале исследуемого периода .

I like it. We trade in astronomical time, not tick time. An Expert Advisor built in time proportional to the trade number can show a nearly ideal straight line, although it will be "not so good", to put it mildly: half of the profit is made at the beginning, on the first 100 trades and in the first year, while the other half is made on the last 100, but in five years (also a straight line, with the same slope, because there are approximately the same amount of successful trades). Let's think about formalization, something like dependence of the system's relative profit on the time interval itself.

Of course, there is no and cannot be a single optimization criterion. Well, formally it is possible to make such a criterion out of dozens of various criteria, but what is the use of it?

 
Mathemat >>:


Единого критерия, конечно, нет и не может быть. Ну формально такой можно составить из пары десятков разнородных критериев, но какой от него толк?

Mathemat, may I ask your opinion - what threshold should be set for the parameters: number of trades, expected payoff and profitability in, say, optimization for a year before you worry about filtering by integral index? For example, after rejecting the results: trades < 50, expected payoff < 50 and profitability < 2, I have no problem choosing among thousands of results because I have either random fly-ins or cluster, but now it's called cloud. From the cloud, I let the machine choose who has the highest Profit * Profitability / Drawdown in %.

I am interested in your expert judgement on the number of deals, expected payoff and profitability during optimisation for the year.

 

Vita писал(а) >> Я, к примеру, после отметания результатов: сделок < 50; матожидание < 50 и прибыльность < 2 не испытываю пробелемы выбора среди тысячи результатов, т.к. остаются либо случайные залетные, либо кластер, но нынче модно говорить облако. Из облака я для себя позволил автомату выбирать у кого больше Прибыль * Прибыльность / Просадка в %.

Vita, OK in principle. Pre-screening is by the number of deals, m.o.s. and PF, and the final screening is by a pretty decent integral criterion along with "cloudiness". That is, in fact, the whole procedure involves filtering by about five different criteria.

The pre-screening figures themselves (m.o., hopefully in the old full points, i.e. at four digits?) are quite logical. I would probably increase the minimum number of transactions (say, to 200) to increase the statistical validity of the results.

 

By the way,this article is interesting. I like the theory of regression, it may be worth implementing...

Profit * Profitability / Drawdown in %. it's good, but for the testing period not to affect the profit index it is better to replace the profit percentage per day (for a lot growing proportionally to the balance), or profit index per day (for a fixed lot). If you don't know how to calculate the percentage per day, here's the formula:

Pr - percentage per day/per trade

Days_Sdelki - number of days or transactions (depending on purpose, percentage of transaction, or percentage per day)

Bal_begin - balance at the beginning of period

Bal_end - balance at the end of the period

Pr=(MathExp(((1/Days_Sdelki)*(MathLog(Bal_end/Bal_begin))))-1)*100;

 

or here is the function


double Procent(double Days_Sdelki,double Bal_begin,double Bal_end)
{
if(Days_Sdelki>1 && Bal_begin!=0) return((MathExp(((1/Days_Sdelki)*(MathLog(Bal_end/Bal_begin))))-1)*100);
else return(0);
}

 

I have introduced one useful variable and am testing a new version of my formula:


PipBar - Pips/Bars (sum of pips for all trades divided by the amount of bars used)

PF - Profit Factor

SdDay - number of deals per day

ProcDay - profit percentage per day (complex formula with logarithms)

MD - Maximum drawdown

SrD - Average drawdown (sum of drawdowns of each order divided by the number of orders)


If(PF>3), Vigoda=2*SdDay+(PipBar/10)+(10*(ProcDay/((MD+SrD)/10));
else Vigoda=(PF-1)*SdDay+(PipBar/10)+(10*(ProcDay/((MD+SrD)/10));


It's a test variant so far, but I'm already happy with the results...

12345678910
New comment