How do I identify the patterns on which the ready-made TS shows a profit? - page 5

 
Alexey Burnakov:
If it's that complicated, then you need to study the price behaviour preceding the trades, enter some price parameters and do a detailed analysis. But it may be difficult to do it in the terminal.
Statistical studies on the level of dependence of parameters, of course, can be done in the same R. It's not about tools, it's about approaches to explaining the workability of actually randomizing one's TS.
 
Alexey Burnakov:
If it's that complicated, you need to study the price behaviour preceding the trades, enter some price parameters and do a detailed analysis. But it might be difficult to do it in the terminal.
We tried it. To make a long story short: for trend strategies the price is moving evenly to the entry point up to the entry point. It is vice versa for contrend strategies. After entering, the position is in a 50/50 condition at any point in time.
 
Vasiliy Sokolov:
Throws. I would be very cautious about such filters.

An element of fitting is always present, for sure. This is why it is particularly silly to throw out the same Tuesdays from a single run. This should be done at least for further optimisation.

I had a long-running case in real life where one day of the week was thrown out. And on each such day I was surprised to see that it was justified by ignoring the day, because the market was losing almost all the time on this particular day of the week. I do not use this filter for a long time, because I do not encounter such a clear difference by days of the week anymore. But I remember well that there may be a pattern that holds good only on certain days. Why that is, I don't know. But it's a fact proven in real life with thousands of trades and many months of trading.

 
zaskok3:
Statistical studies on the level of parameter dependencies can of course be done in the same R. It's not about tools, it's about approaches to explaining the performance of your TS's actual randomness.

The most straightforward way is to compare the financial result of a trade using the TS with randomly generated results. Do a monte carlo experiment, generating n times (10,000 times for example) the result of a random TS, keeping the number of trades = degree of freedom. The duration of the trades and the buy/sell percentage should also be identical.

Using the obtained random results take the 99.9%-quantile. If your TS exceeds it, we can say that at the level of significance 0.01 or less your TS is exploiting some profitable relations.

 
Alexey Burnakov:

The most straightforward way is to compare the financial result of a trade using the TS with randomly generated results. Do a monte carlo experiment, generating n times (10,000 times for example) the result of a random TS, keeping the number of trades = degree of freedom. The duration of the trades and the buy/sell percentage should also be identical.

Using the obtained random results take the 99.9%-quantile. If your TS exceeds it, we can say that your TS exploits some profitable relations at a significance level 0.01 or less.

I'm aware of the monte carlo. And have seen you do it in particular. But there is no faith in this method. Suffice it to say that it generates a price series by mixing deductions. And it's the deductions that are the weak point. Because they can be calculated in different ways. For some reason, deductions are calculated using timeframes, which are completely artificial fictitious entities. How to calculate deductions correctly is a big question. And is it necessary to compute them to make artificial series, or should they be generated by a different method than mixing in general? Extremely ambiguous.

There is no doubt that there is a pattern. The question is, what is the pattern?

Теория информации в задаче проверки гипотезы о независимости значений, принимаемых случайной переменной, на примере индекса DJI
Теория информации в задаче проверки гипотезы о независимости значений, принимаемых случайной переменной, на примере индекса DJI
  • habrahabr.ru
Только зарегистрированные пользователи могут оставлять комментарии. Войдите, пожалуйста. Пометьте топик понятными вам метками, если хотите или закрыть
 
zaskok3:

I know about Monte Carlo. And have seen you do it, in particular. But there is no faith in this method. Suffice it to say that it generates a price series by mixing deductions. And it's the deductions that are the weak point. Because they can be calculated in different ways. For some reason, deductions are calculated using timeframes, which are completely artificial fictitious entities. How to calculate deductions correctly is a big question. And do they need to be computed to make artificial series, or should they be generated by a different method than stirring at all? Extremely ambiguous.

And there is no doubt that the pattern exists. The question is - what is the pattern?

That example doesn't really fit the context. I'm not suggesting mixing increments in order to race TC on them. I'm suggesting to make an algorithm where all important parameters of the TS will be randomly selected each time, but the work will go on the original price series.

Although, it's a slippery slope. If the TS is fitted to the price series, we can generate the same fitted TS.

In general, if we want to directly understand the patterns, we should investigate the price along with ongoing deals (their results), but because of the vagueness of the TS, it will be difficult to find something.

 
Alexey Burnakov:
I'm not suggesting mixing increments to race the TS on them. I am proposing to make an algorithm where all important parameters of the TS will be randomly selected each time, but the work will go on the original price series.

This is how I do bruteforcing when I optimise. Accordingly, all results for any set of input parameters are available. But how will this answer the:

zaskok3:

Question - what is the pattern?

 
zaskok3:

This is how I do bruteforcing when I optimise. Accordingly, all results for any set of input parameters are available. But how does this allow for an answer to:

Answered:

It's a slippery slope though. If the TS is fitted to a price range, then you can generate similarly fitted TSs.

In general, if you want to understand the patterns directly, you need to investigate the price in conjunction with the trades taking place (their results), but due to the fuzziness of the TS itself, finding something will be difficult.

 
zaskok3:


There is no official list of patterns and their names. Or maybe there is, but I haven't seen it. I come up with a pattern myself, then only write the entry condition according to this pattern. Only after the test it will be clear whether it's a real pattern or my fantasy. For example: "If the price has passed 20 points within 5 bars, then it will go further by at least 5 points". Suppose it all worked. What should we name this pattern? The entering condition is its name. The Expert Advisor can have many conditions for entering, but there is the main condition. If we remove it, the order will not open. This condition is the name of the pattern or this condition + something else.
 
David Azizian:
There is no official list of regularities and their names. Or maybe there is, but I haven't seen it. I invent the pattern myself and then I only write the condition for entering by this pattern. Only after the test it will be clear whether it is a real regularity or my fantasy. For example: "If the price has passed 20 points within 5 bars, then it will go further by at least 5 points". Suppose it all worked. What should we name this pattern? The entering condition is its name. The Expert Advisor can have many conditions for entering, but there is the main condition. If we remove it, the order will not open. This condition is the name of the pattern or this condition + something else.

Agreed!

I still don't understand the author. Either he has a black box sewn into the trading algorithm and is trained with so many scales that it is impossible to understand the rules of the box itself. Or the rules of trading are clearly prescribed in the code, but he wants to invent names for market regularities.

For me, I will first form the rules of entering and exiting a position and this will already be the identified patterns: If, for example, the price went down by 80 points and then went up by 10 points, it will go back down and hit the TP of 50 points with the probability of so-and-so.