Machine learning in trading: theory, models, practice and algo-trading - page 3302

New comment
 
Anyone have any thoughts on why a spread of 2 five digit points kills alpha on eurobucks five minutes? I mean, without it, it's a grail, with it it's bad. But it has almost no effect on hourly trades.

Moreover, if you run a 5-minute bot on hourly markets, it works. The only difference is opening on a new bar every hour.

I subtracted the spread from the increments, visually not much has changed. It looks more like the markup of trades does not take it into account properly.

 
Maxim Dmitrievsky grail, with it it's bad. But it has almost no effect on hourly trades.

Moreover, if you run a 5-minute bot on hourly markets, it works. The only difference is the opening on a new bar every hour.
.

I subtracted the spread from the increments, visually not much has changed. It looks more like the markup of trades does not take it into account properly.

Version 1
probably gets a lot of entries at about 1 price level (on flat), which are then drained.

Version 2
Trading 12 times more often (60/5) you reach a drain faster. Test H1 on the same number of signals you received from M5.

 
Forester #:

Version 1
probably get a lot of inputs at about 1 price level (on flat), which then merge

Version 2
Trading 12 times more often (60/5) you reach the drain faster. Test H1 on the same number of signals that you received from M5

So far I stick to the version that cloze prices of five minutes are close to each other, so the spread eats the profit. Now afk, then I will check different options .

It is possible to make at least closing not on a new bar, but on ticks. The funny thing is that I put the spread into the model at the training stage. That is, to make trades strictly greater than the spread. And still it happens.

I also subtracted the spread from the increments, and trained on the resulting less volatile, by the spread value, chart. Then I tested it on the original one. It seems to correct the picture a bit, but still not the same.
 
Published articleScikit-learn library classification models and their export to ONNX
 
Maxim Dmitrievsky #:
So far, I'm sticking to the theory that the cloze prices of the five minutes are close together, so the spread eats up the profit.
I think it is more complicated. For example, there is an Expert Advisor in Market that shows striking results on real ticks and tester-generated ticks. It is not possible to get into the reasons.
 
Maxim Dmitrievsky grail, with it it's bad. But it has almost no effect on hourly trades.

Moreover, if you run a 5-minute bot on hourly markets, it works. The only difference is the opening on a new bar every hour.
.

I subtracted the spread from the increments, visually not much has changed. It looks more like the markup of trades does not take it into account properly.

Calculate the quantile of price increments with a small step, for example, 0.01. You will probably see that in hourly trades the increments are larger than the spread in hourly trades above the spread is many times larger than on M5.

 
fxsaber #:
I think it is more complicated. For example, there is an Expert Advisor in the Market that shows different results on real ticks and tester-generated ticks. I can't get into the reasons.
Probably it is based on non-true optics
 
СанСаныч Фоменко #:

Calculate the quantile of price increments with a small step, for example, 0.01. You will probably see that in hourly rates the increments are greater than the spread in hourly rates above the spread is many times greater than on M5.

Well, it is logical, but trades are placed with the spread taken into account. The signal can be many bars ahead. I increase the spread - it becomes worse and worse even on the traine, that is a direct dependence.
 
Maxim Dmitrievsky #:
Probably on the OOS.

On OOS. I'll text you.

 
fxsaber #:
I think it is more complicated. For example, there is an Expert Advisor in the Market that shows different results on real ticks and tester-generated ticks. It is not possible to get into the reasons.

Download the Expert Advisor, disallow updates, check it in a week on fresh ticks.

Or shift the time in the history, there must be a catch.

1...329532963297329832993300330133023303330433053306330733083309...3548
New comment
Reason: