Is there a pattern to the chaos? Let's try to find it! Machine learning on the example of a specific sample. - page 8

New comment
 
Aleksey Vyazmikin #:

Yes, there is a lot of data, and I plan to add more - so developing a methodology for screening before obfuscation is required.

Just now I am training, and I will say that a lot depends on the settings, especially the number of splits in the quantum tables.

I've just started an experiment where training is done with default settings on the video card - one pass, without taking into account model evaluation and test on exam sample, takes 2-3 minutes - depending on the resulting number of trees in the model. On my rather outdated FX-8350 processor, it is about 60% slower.

I think the speed is quite acceptable, I usually train 100 models with fixed seed to average the efficiency of the method.

If you train to the "end", the programme estimates the time up to 2 hours.

2-3 minutes with a tree depth of 6 and 1000 trees?
 
spiderman8811 #:
Not levels, but ranges plus there wave patterns and on candlesticks. Those aren't in the books. It should work.
I am also interested in more details)))) it is not clear from what range and what model).
 
elibrarius #:
2-3 minutes with a tree depth of 6 and 1000 trees?

Trees 250-400 are built, as there is a control of stopping training on the test sample, i.e. if there is no improvement during the last 100 trees, the training stops and the model is cut to the last tree with improvement.

 
Aleksey Vyazmikin #:

Here's another variant - I like it even better, as it's a stable result on all samples.



0.042 is the best result. Better than on all columns and the balance curve is prettier. But not as good as you did on Catbusta.


 
Aleksey Vyazmikin #:

Is the first column zero or "1"? :)

That's zero.

0,02400


 
Aleksey Vyazmikin #:

Probably close in meaning to 1041-1489.

At 448 bars, the best 0.03000


 
elibrarius #:

0.042 is the best result. Better than all the columns and the balance curve is prettier. But not as good as you did on Catbusta.


elibrarius #:

At 448 bars the best is 0.03000.


The result is already clearly better, and it seems to have been achieved because of the choice of learning-enhancing predictors. How many other useful ones there are and how to get them out is the concern.

Try to change the target by making "1" only if you achieve a profit of more than 50 pips (maybe less is better) - this improved learning in my experiments, although the number of positive targets got even smaller...

 
Aleksey Vyazmikin #:

The result is already clearly better, and it seems to be due to the choice of predictors that favour learning. How many more of them are useful and how to get them out is what is of concern.

Try changing the target by making "1" only if profit is achieved over 50 pips (maybe even less is better) - this improved learning in my experiments, although the number of positive targets got even smaller...

The 2nd column is the class threshold (but not in the teacher's markup, but in the forecast). 3rd profit.
The graph for the 60 pt porgoa is the best.
 
elibrarius #:
The 2nd column is the class threshold (but not in the teacher's markup, but in the forecast). The 3rd is profit.
The graph for 60 pts is the best.

And how do you know the profit when forecasting, or do you have a regression model?

Try shifting when teaching :)

 
Aleksey Vyazmikin #:

And how in forecasting do you know the profit, or do you have a regression model?

Try shifting when training :)

I calculate profit and schedule after training.
123456789101112131415...32
New comment