Machine learning in trading: theory, models, practice and algo-trading - page 2009

 
Aleksey Vyazmikin:

Where to get these "normal" ones!? And it affects or not - here is random, if during quantization we hit the boundary between the quantiles, then during training it may turn out that just the difference of 1 quantile will lead to missing an entry on the real. I do not like it when I cannot reproduce a closed day's trades in the Strategy Tester; that is why I detected the reason and started to eliminate it.

If we take a hypothetical normal TS) it would be foolish to think that +-lag would change something
 
Maxim Dmitrievsky:
If we take a hypothetical normal TS) it would be foolish to believe that +-lag would change something

Well, how silly, let's say the lag (the obtained wrong value) means the crossing of an important level by the price - and statistically - there is a crossing, it means a reversal soon and it makes no sense to follow the trend on this hourly bar, but in fact there was no crossing, but only a touch (the bar is not closed), so there is still a potential for further movement at the bar opening.

And then, as I said before, it is important for debugging and analysis to be able to reproduce the results of the trading in the tester.

 
Aleksey Vyazmikin:

Well how silly, let's say - lag (obtained incorrect value) means a crossing by the price of an important level - and statistically - there is a crossing, it means a reversal soon and it makes no sense to enter the trend on this hourly bar, but in fact there was no crossing, but only a touch (the bar is not closed), so there is still potential for further movement at the bar opening.

And then, as I've already mentioned, it is important to be able to reproduce the results of the trading in the tester for debugging and analysis.

The margin of error is insignificant. If there is a shift or change of series in 10% of sample then yes, while a change of less than one percent should not change the result more than one percent.
 
Valeriy Yastremskiy:
The error rate is insignificant. If there is a 10% sample shift or change in the row, then yes, but a change of less than one percent should not change the result by more than one percent.

Strange how you think. if it's a single tree, then the change can be critical when the error goes beyond the split threshold. If it's a scaffold or, boosting, then over the course of the model the error will be critical in a couple of "predictions", which is not fatal. But, this is if we are talking about one predictor, but if there are 30% of such predictors, the results will differ more often.

 
What is the correct way to do something only once on the bar opening inside OnTick()?
I use this crutch
if(BarTime!=iTime(NULL,PERIOD_M3,0)
{
 BarTime=iTime(NULL,PERIOD_M3,0);
 ...
}
There must be some nice solution.
 
Evgeny Dyuka:
What is the right way to do something only once on bar opening inside OnTick()?
I use this crutch
there must be some nice solution.

Classic)

 
Aleksey Vyazmikin:

It's a strange way of thinking. if it's a single tree, then the change may be critical when the error goes beyond the split threshold. If it's a scaffold or, boosting, then over the course of the model the error will be critical in a couple of "predictors", which is not fatal. But, this is if we're talking about one predictor, and if such predictors 30%, the results will already differ more often.

Well it's just to the question of the relative amount of uncorrected data. Of course there is a critical threshold.

 
Valeriy Yastremskiy:

Well it is just to the question of the relative amount of incorrect data. Naturally, there is a critical threshold.

So it will be, if the data are taken at different time and their request will be on the bar index from different instruments especially.

 
Can we have a contest to learn by sampling?
 
Aleksey Vyazmikin:
Can we hold a contest for sampling training?
Sampling needs to be formalized somehow. Otherwise, suddenly the phasers will be brought in and win.)