Is there a pattern to the chaos? Let's try to find it! Machine learning on the example of a specific sample. - page 10

 

train

This is how the model looks on the training sample - you can see a good delta margin between the green and red curve - this is profit.

But below we can see how the delta has shrunk on the exam sample

Compared to the tester, the calculated balance turned out to be a bit more optimistic, but the structure is identical - I think I will continue to use it for initial evaluation.


 
Aleksey Vyazmikin #:

This is how the model looks in the training sample - you can see a good margin of delta between the green and red curve - this is profit.

But below we can see how the delta has shrunk on the exam sample

Compared to the tester, the calculated balance turned out to be a bit more optimistic, but the structure is identical - I think I will continue to use it for initial evaluation.


0.10500 is the best option. It's about the same as yours. But the balance line is different. And the error is about 0.5. It's risky, it will get a little worse and it might start to drain. 4200 trades and only 0.10500 /4200 ~= 0.00002 per trade. Very risky. Spread, slippages, etc. will eat up all the winnings.


 
elibrarius #:
0.01050 is best. It's about the same as yours. But the balance line is different. And the error is about 0.5. It's risky, it will deteriorate a little bit and may start to drain. 4200 trades and win only 0.01050 /4200 ~= 0.00002 per trade. Very risky. Spread, slippage, etc. will eat up all the winnings.


Due to the model the percentage of profitable trades increased by 4%, plus MM will give the same amount - and now you can think about exploitation.

But I think that this markup is not correct, because it is not based on the market structure - there is no attempt to compare similar market conditions for training, so the model has to do everything by itself.

 
Also, I think that the balance should be determined in the end by two models (buying and selling) - after all, they can self-compensate.
 
Aleksey Vyazmikin #:
Also, I think that the balance should be determined in the end by two models (buy and sell) - because they can self-compensate.
I agree, that's what I do for my experiments, different classes should not interfere with each other when training. 1 model will look for the overall best result. The 2 best models should be overall better than one. But on the other hand they can overtrain faster, i.e. overtraining should be blocked more strongly.
 
Aleksey Vyazmikin #:
Also, I think that the balance should be determined in the end by two models (buy and sell) - because they can self-compensate.
Learn on the first 2 columns) On the last sample on H1.
 
elibrarius #:
Train on the first 2 columns) On the last sample on H1.

Does the time pattern pick up?

 
Aleksey Vyazmikin #:

Is it picking up a temporal pattern?

I do. See what you get
 
elibrarius #:
I do. See what you get

I'm having a bit of fun with a different approach at the moment - no chance to test it yet. But I think it will also find it, if it's obvious.

 
Aleksey Vyazmikin #:

I'm having a bit of fun with a different approach now - no opportunity to check it yet. But I think it will also find it, if everything is obvious there.

The point is that it is 2 times better than on 5000+ features.
It turns out that all the other 5000+ features only worsen the result. Though if you select them, you will surely find some that improve.
It is interesting to compare what your model will show on these 2.