Machine learning in trading: theory, models, practice and algo-trading - page 1535

[Forester]

Maxim Dmitrievsky:

increments

I am going to finish the connector this weekend and put it to the tests. I plan to upload the model to the cloud and use the terminal to retrieve signals. I have a good idea how to use it in the future. I can send the client on MT5 later

I would like to use the demo for a couple of months.
And then I would like to start copying or repeating the method.
My experiments with Alglib forest and single tree with input price increments and target TP/SL - didn't show anything interesting.

[Maxim Dmitrievsky]

elibrarius:
For starters, it's interesting to just look at the demo for a couple of months.
And then invest in copying or repeating the method.
My experiments on alglib forest and single tree with input price increments and target marked by TP/SL - didn't show anything interesting.

I got to the correct design of the targets only after a year, and by changing their parameters you can change acurasi in a predictable way

on the demo yes, the tests will show

alglib forest has a problem that I don't understand - the more samples it has, the more it overtrains

Maybe it makes sense to do van hot for it, i.e. convert features into categorical ones. Trees will not be so huge

[Maxim Dmitrievsky]

Maxim Dmitrievsky:

alglib forest has a problem that I don't understand - the more samples it has, the more it overtrains

Maybe it makes sense to do a van hot for it, i.e. to convert features into categorical ones. Trees would not be so huge

on the subject

Some people write that vanchot is bad for scaffolding because it makes trees asymmetric, others write that it prevents overfooting. Chose the second one :)

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6368971/

[Forester]

Maxim Dmitrievsky:

for the sake of argument.

Some write that vanchot is bad for forests because it causes asymmetric trees, others write that it prevents overfitting. Chose the second one :)

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6368971/

You can simply discretize the input data, e.g. convert 5 digits to 4 digits. And the data will already be in groups of 10.
Or, as I suggested earlier, you can build in the alglib forest to stop branching when the desired depth or number of examples in the sheet is reached.

[Maxim Dmitrievsky]

elibrarius:

You can simply discretize the input data, e.g. convert 5 digits to 4 digits. And the data will already be in groups of 10.
Or as I suggested before - to build in alglib forest stopping branching when required depth or quantity of examples in the list is reached.

This is different, you have to make categorical ones so that their values cannot be compared with each other.

I'll get the article later, I've discounted it so I don't forget it

[Aleksey Vyazmikin]

Maxim Dmitrievsky:

Why bother with this... big companies like Yandex work, they do things. They say do it this way and you'll be fine. Just do it this way and don't do amateur work. Otherwise you will drown in wording and different approaches.

It shows the change as it goes along the gradient, building up trees

There changes go when adding a tree, right, while I need to look at the sample, as if adding a new row to the sample - probably so you can understand which areas/situations turn out to be difficult to learn, and accordingly think about predictors that can overcome these difficult areas to learn.

[Aleksey Vyazmikin]

elibrarius:
Not bad! Increase the volume and something can already be earned)

So to increase the volume, you select leaves and combine them into one model, which is how you increase the number of inputs, i.e. Recall.

Perhaps with your sample this method will work, but if it were accelerated...

Such results are given by tree with depth of 5-7 splits, but it reveals a certain pattern only - usually 3-4 sheets go to -1 and 1, and the rest to zero.

[Aleksey Vyazmikin]

Maxim Dmitrievsky:

increments

How many predictors are there in total?

[Aleksey Vyazmikin]

elibrarius:

You can simply discretize the input data, for example convert 5 digits to 4 digits. And the data will already be in groups of 10.
Or, as I suggested earlier, you can build a branching stop into alglib forest when the desired depth or number of examples in the list is reached.

Of course, you should pruning or simply prohibit branching by the number of examples (completeness - Recall) in a single leaf. In addition, you can reject trees if their accuracy and completeness values are below the plinth.

[Aleksey Vyazmikin]

Maxim Dmitrievsky:

This is different, you need to make categorical so that their values can not be compared with each other

I'll get the article later, I threw it away so as not to forget

By the way, what are your parameters for model training?