Optimisation! Share your experiences, please. - page 5

 
i have a good idea, i'll show the drawdown for these two weeks more than i expected from Krafik for the last year ? (I mean it may be normal drawdown if I just wait a month and the chart will straighten out) and another question, what will this EA show if I re-timed it for 2006 and then reset it for 3 months 2007 ? ( I have a lot of experience in testing my EAs and good ones are more or less stable for more than a year ) <br / translate="no">
The thing is that I have been developing this EA during the whole year of 2006. I have found about 5 versions of settings. 3 of them died since the beginning of 2007 - i.e. drawdown is more than for the period of 2006. Two of them are still alive, but not the same dynamics - not coinciding with 2006, but they are not draining. As for stability in different years - I can't boast. 2005 had half year growth, then slump, and then recovery. I.e. a big dip of about 60%. In 2004 all seems to be not bad - i.e. there are no plums, but yield curve is "kasturbata" (sorry for the expression). The variant of the test that is presented here is the Expert Advisor reoptimized using the data till the 16th of February and has been tested almost at the current moment. There is an Expert Advisor for Day Boards with similar logic, there is 3-4 years of growth, then flat on the yield curve for a couple of years, then growth again. In general, everything is not clear. I am afraid of getting the EA fitted via optimization. I don't know how to determine what may be allowed to go real :(

My opinion :
Need to refine the algorithm or look for a new one .... I would be afraid to bet on a real account in such a situation... i'd be wary of betting on a real... how do i know it won't happen again in 2005... you can visually analyze why it failed in 2005 and optimize the algorithm (it may be useful) you don't have many deals .... With so many trades, I would even optimize it in 7 years. it would be hard to fit :))
IMHO
 
I try to put things on the real for which you can say "EVERYTHING is doubtful".
 
nchnch:

my opinion :
We need to refine the algorithm or look for a new one .... I'd be afraid to bet on the real in this scenario... where's the guarantee that the 2005 situation won't repeat itself... you can visually analyze why it failed in 2005 and optimize the algorithm (it may be useful) you don't have many deals .... With so many trades, I would even optimize it in 7 years. it would be hard to fit :))
IMHO

I tried... But the thing is, if the strategy is successful in principle, then the year differs from year to year only in parameters.
With one set, the strategy rises in 2005, but loses in 2006.
The other set, on the contrary, rises in 2006, and in the tank-test 2005 fails.
The 2007 one is still going up, as soon as it starts to turn strongly we will optimise it!
 
nchnch:
i have a good idea, i'll show the drawdown for these two weeks more than i expected from Krafik for the last year ? (I mean it may be normal drawdown or it may be just wait a month and the graph will level out) and another question, what will this EA show if it was optimized for 2006 and then reset for 3 months 2007? ( I'm very experienced in testing my EAs and good ones are more or less stable for more than a year)

The thing is that the data period we used for testing was the whole year 2006. I have found about 5 variants of settings. 3 of them died since the beginning of 2007 - i.e. drawdown is larger than in the period of 2006. Two of them are still alive, but not the same dynamics - not coinciding with 2006, but they are not plummeting. As for stability in different years - I can't boast. 2005 had half year growth, then a plum, and then a recovery. I.e. a big dip of about 60%. In 2004 all seems to be not bad - i.e. there are no plums, but yield curve is "kasturbata" (sorry for the expression). The variant of the test that is presented here is the Expert Advisor reoptimized using the data till the 16th of February and has been tested almost at the current moment. There is an EA with similar logic for days with 3-4 years of growth, then it goes flat on the yield curve for a couple of years, then it grows again. In general, everything is not clear. I am afraid of getting the EA fitted via optimization. I don't know how to determine what I am allowed to do in real trading :(

My opinion :
Need to refine the algorithm or look for a new one .... I would be afraid to bet on a real account in such a situation... i'd be wary of betting on a real... how do i know it won't happen again in 2005... you can visually analyze why it failed in 2005 and optimize the algorithm (it may be useful) you don't have many deals .... With such a large number of trades, I would even optimize it in 7 years. it would be hard to fit :))
IMHO

OK, my advice - optimize in 7 years!
Accepted:) Will be straining the computer brain now.
My asya 15455524. If interested in the result, knock. Or in general, to discuss a sore subject :)
 
I also have a question about optimization:)

I optimize only by Take and Stop Loss. Can we trust this optimization? In the tester everything is nearly the same - one year down, another one up (i.e. parameters TP and SL adjusted for one year do not show the same results in the next year). Parameters of the Expert Advisor are not changed and chosen "by eye". So the question arises: Can we "overoptimize" the Expert Advisor, changing only TP and SL?
 
sashken:
I also have a question about optimization:)

I optimize only by Take and Stop Loss. Is it trustworthy? In the tester it is about the same - one year down, another one up (i.e. parameters of TP and SL adjusted for one year do not show the same results in the next year). Parameters of the EA do not change and are selected "by eye". So the question arises: Can the EA be "re-optimized" by changing only the TP and SL.

I trusted it. So far I'm in the black (7 months).
 
PSmith:
sashken:
I too had a 'born' question about optimisation:)

I trusted it. So far I'm on the plus side (7 months).

Thanks for the reply, I also think that it seems to be possible to trust.
 
sashken:
I also have a "born" question about optimisation:)

The strategy inherent in my Expert Advisor works in principle. I optimize only by Take and Stop Loss. Is it trustworthy? In the tester it is about the same - one year down, another one up (i.e. parameters of TP and SL adjusted for one year do not show the same results in the next year). Parameters of the EA do not change and are selected "by eye". So the question arises: Can the Expert Advisor be "reoptimized" by changing only TA and SL?
At present time the best results are shown only for those optimization parameters whose yield curve is upward (it filters itself out if we disable useless results) and linear correlation coefficient of this very curve is closer to 1 in absolute value. I.e. the program takes one by one variants given by the optimizer, runs a test on each of them and analyzing profits and losses finds the one having the graph most similar to a straight line. Obviously, such a graph will never give the best balance of the optimization result, in most cases it is a middle one. But profitability curves that are more linear have a very small drawdown and abnormally sharp upward moves are also not observed.

Now I think to stop using this optimization method as it takes too much time, i.e. for each result I need to run the tester and run it, and then calculate the regression on the yield curves. It's just that each run of the terminal from the command line for testing is too slow and waiting a long time.

So far I have thought that after each optimization run I should take all trades history in deinit() event and directly count correlation and output the results, i.e. parameters of Expert Advisor and correlation coefficient into csv format file. Then, after optimization, I only need to sort the same file by the coefficient and feed the necessary parameters to my Expert Advisor. I am currently finalizing this variant.

It would be nice if the developers had added the linear correlation coefficient by the yield curve in the tester and asked to sort the optimization results by this very coefficient. But it is easier to do it myself until you ask them.
 
If the Expert Advisor opens positions correctly using a workable TS - the optimization curve should not have strong dips and spikes, and if they happen, you should analyze the cause, it is best to take average, but stable results.
 
Reshetov:
sashken:
I also have a "born" question about optimization:)

I optimize only by Take and Stop Loss. Is it trustworthy? In the tester it is about the same - one year down, another one up (i.e. parameters of TP and SL adjusted for one year do not show the same results in the next year). Parameters of the EA do not change and are selected "by eye". This raises the question: can the Expert Advisor be "reoptimized" by changing only the TP and SL.
At the moment the best results are shown with optimization parameters having yield curve upwards (it filters itself out if we disable useless results) and linear regression coefficient of this very curve is closer to 1 in absolute value. I.e. the program takes one by one variants given by the optimizer, runs a test on each of them and analyzing profits and losses finds the one having the graph most similar to a straight line. Obviously, such a graph will never give the best balance of the optimization result, in most cases it is a middle one. But profitability curves that are more linear have a very small drawdown and abnormally sharp upward moves are also not observed.

Now I think to stop using this optimization method as it takes too much time, i.e. for each result I need to run the tester and run it, and then calculate the regression on the yield curves. It's just that each run of the terminal from the command line for testing is too slow and waiting a long time.

So far I have suggested that after each optimization run the deinit() event should take the entire history of deals and directly calculate the regression and output the results, i.e. parameters of the Expert Advisor and regression coefficient into csv format file. Then, all that remains after the optimization is to sort the file by the coefficient and load the necessary parameters into the Expert Advisor. I am currently developing this variant.

It would be nice if the developers would also insert the coefficient of linear regression along the yield curve in the tester and allow to sort the optimization results by this very coefficient. But one has to ask them to do it themselves.
Great! I agree. The yield curve should be straight. Thanks for the idea. I also want to do so, but I'm afraid I don't have enough experience to program it :(