Machine learning in trading: theory, models, practice and algo-trading - page 3439

New comment
 
Maxim Dmitrievsky #:

it's easy to calculate stats and show prufs

but I don't mind the opens )

I'm in the middle of something else right now,

prufs is any commercial tester.

prufs is that we get the signal at the close of the candle, which means that we can't open a position before the open.

The proof is that a slose is not equal to an open and it's an unaccounted cost.

 
mytarmailS #:

I'm in the middle of something else right now,

the prufs are any commercial tester.

The prufs is that we get the signal at the close of the candle, which means that we can't open the position before the open.

The proof is that the slose is not equal to the open and it is an unaccounted cost.

if a 1 tick difference breaks the TS, then it's a bad TS ) or a crooked toaster.

 
Maxim Dmitrievsky #:

If a 1 tick difference breaks the TC, it's a bad TC ) or a crooked toaster

The toaster is honest, compared and checked with commercial the whole book of deals, not just by eye on the capital curve.

 
mytarmailS #:

Toaster is honest, compared and checked with commercial the whole book of transactions, not just by eye on the capital curve.

I don't know, I'll compare it later.

 
Rorschach #:

Even more. On n1 out of all possible combinations of TC parameters, plus show ~60%, on m1 already under 90.

I also trained on one pair and tested on other majors. All the same it gave a plus. Only yen should be excluded, it spoils all the results.

What if you remove increments from the sample that do not cover the specified spread.

Or some other filtering methods.

 

Create a custom symbol.

Tick.bid = (Tick.bid + Tick.ask) / 2;
Tick.ask = Tick.bid;

Then in the Tester gradually increase the spread and see how the results change.


 
Maxim Dmitrievsky #:

If a 1 tick difference breaks the TC, it's a bad TC ) or a crooked toaster

You have shown yourself that the spread breaks the TS as well. 1 tick (between Close and Open including), probably 2-3 pips on average. Spread on ECN on EuroBucks is about the same. So it is possible...
But this is not all the problems. Even if you send a trade command to Open, it will still be executed on the next tick at best, and if your computer is far away from the DC, you will get another 20-200 ms of delay (and if you send it to a liquidity provider (for large volumes), even 2-5 sec), which can add a few more ticks.

You can trade limits like FXsaber, but probably 30-50% will not work, because the price may not return to the set price of the limit. I.e. the model was taught as if all of them were executed, but in reality not all of them will be (and it is possible that the best ones will not be executed when the price rushed up and did not return, i.e. strong movements will be missed).
.
fxsaber #:
Do you have any statistics, how many limits remain unexecuted on average?
How much time should be given for execution? It seems to me that after 1-2 minutes we should take them off, because the situation will change.
 
Forester #:
You have shown yourself that the spread breaks the TS. 1 tick (between Close and Open including), probably 2-3 pips on average. Spread on ECN on EuroBucks is about the same. So it is possible...
But this is not all the problems. Even if you send a trade command to Open, it will still be executed on the next tick at best, and if your computer is far away from the DC, you will get another 20-200 ms of delay (and if you send it to a liquidity provider (for large volumes), even 2-5 sec), which can add a few more ticks.

You can trade limits like FXsaber, but probably 30-50% will not work, because the price may not return to the set price of the limit. I.e. the model was taught as if all of them were executed, but in real life not all of them will be (and it is possible that the best ones will not be executed when the price rushed up and did not return, i.e. strong movements will be missed).
.

Yes, they will not be executed at all. Only bad ones will be executed.

That is why I always strongly discourage trading on small TFs.
 
fxsaber #:

Create a custom symbol.

Then in the Tester gradually increase the spread and watch how the results change.

I do it in my toaster, no problems with it. The problem is how to make it work when there is a spread :)

 

2. Cost of cloud resources

Using cloud resources for model training can be costly, especially if the training process takes a long time. Therefore, it is advisable to optimise the use of these resources, e.g. by pre-processing data locally or using cloud resources only for training the most complex models.

This is a point that can be solved somehow and the training speed will increase!

1...343234333434343534363437343834393440344134423443344434453446...3584
New comment