Machine learning in trading: theory, models, practice and algo-trading - page 3116

New comment
 
Evgeni Gavrilovi #:

And if we take into account the increments of other currency pairs? Would it be useful?

The mere fact of adding increments of other symbols will not improve the results if the approach does not work without them. If it works without them, there may be variants. Either if the bot is planned to be multicurrency. Or if you want to deduce something common between different quotes, but it will be some other approach.
 
Evgeni Gavrilovi #:

And if we take into account the increments of other currency pairs? Would it be useful?

"Impact" is a qualitative characteristic.

You need a quantitative measure of the strength of the relationship between the predictor and the target. I have written many times on this forum, made references to R packages, even cited the results of my calculations.

The very idea to include some predictors based on other currency pairs is quite working.


PS. If you do not use such a measure of linkage at the preprocessing stage, you should not talk about MO at all.

 

Contrary to the outdated information that preprocessing is "our everything", both preprocessing and postprocessing have been well researched and proven to be effective.

For example, feature learning (or representation learning) is related to preprocessing and has proven itself in various tasks.
 

Let us present a hypothetical situation with one theoretical TS, which consists of a basic model that predicts the direction of the trade and a meta model that predicts the probability of winning (to trade or not to trade):

Let's call the first model the main model, which divides the feature space into buy/sell with a black line. And the second is a meta model that divides the total feature space into trade/don't trade (red line).

Now let's imagine another variant, when there are two meta models and each of them divides different feature spaces of BUY and SELL classes into trade/non-trade separately (two red lines).

A purely theoretical question "to think about" is whether the second option is better. And if it is better, why. Please comment.

A request, probably even to Alexei Nikolaev, how one can determine the effect of such "intervention". After all, we will get 2 probability distributions of two meta models, which can be compared/evaluated/distributed by corners.
 

I use the 2nd variant. I have not tried the 1st one, because it immediately caused scepticism.
I think that bulls and bears trade differently. The same euro usually falls quickly and then slowly creeps up. Different behaviour. Different chips can become important too. Different hyperparameters in the models. One on buy/sell is unlikely to combine well the different behaviour of different actions. There will be something in between.

 
Forester #:

I use the 2nd variant. I have not tried the 1st one, because it immediately caused scepticism.
I think that bulls and bears trade differently. The same euro usually falls quickly and then slowly creeps up. Different behaviour. Different chips can become important too. Different hyperparameters in the models. One on buy/sell is unlikely to combine well the different behaviour of different actions. It will be something in between.

Intuitively it seems that way too. But it is also possible to get probabilities to trade/not to trade for buy/sell of two models at once, regardless of which direction is predicted by the main model. Compare them and make an additional check, so that the probabilities differ significantly for opening a trade.
 
Maxim Dmitrievsky direction of the trade and a meta model that predicts the probability of winning (to trade or not to trade):

What does it mean to predict the direction of the trade

What do you mean it predicts the probability of winning?

These are too vague concepts...


In general cases, binary classification of market growth/decline forecast in the form of probability will solve this problem.


the probability of growth is more than 0,5 - this is the direction of the deal.

probability of high, for example 0,8 - it will be the probability of winning.

And don't meta models.

But this is in general cases, but I understand that we are not talking about general cases, so we need to clarify the terminology that is


predicts the direction of the deal

predicts the probability of winning


 
Maxim Dmitrievsky #:
Intuitively, it seems that way too. But it is also possible to get probabilities for buy/sell from two models at once, regardless of which direction is predicted. Compare them and make an additional check, so that the probabilities differ significantly for opening a trade.
I haven't tried it. Intuitively) But as Marx said: practice is the criterion of truth. If it works for you in practice - good)
 
mytarmailS #:

What do you mean, predicts the direction of the transaction

What does it mean - predicts the probability of winning

It's a fuzzy concept.


In general cases binary classification of market growth/decline forecast in the form of probability will solve this problem


growth probability greater than 0.5 is the direction of the trade

probability is high for example 0,8 - this will be the probability of winning

And no meta models.

But that's in general cases, but I understand that we're not talking about general cases, so we need to clarify the terminology of what is.


predicts the direction of the transaction

predicts the probability of winning


This has all been discussed here many times in the prehistory. The first model trained on buy/sell is tested on new data. Cases where it is wrong are put into the class of not to trade, the rest into trade. The second classifier is trained on it. We get two models. One of them predicts the direction, the other one predicts whether the trade should be taken out. This gives flexibility, if we just set a threshold for trades through one model. Because the 2 models can be improved, one through the other. I described the original method in the last article. Then I switched to modified logic.

This, by the way, is an open question, because it is possible to improve each through the other, apparently, in different ways.

Then I came across kozul inference, where they do it in a similar way.

 
Maxim Dmitrievsky #:

This has all been discussed here many times in the prehistory. The first model trained on buy/sell is tested on new data. Cases where it is wrong are put into the class of not to trade, the rest into trade. The second classifier is trained on it. We get two models. One of them predicts the direction, the other one predicts whether the trade should be taken out. This gives flexibility, if we just set the threshold for trades through one model. Because the 2 models can be improved, one through the other. I described the original method in the last article. Then I switched to a modified logic.

This, by the way, is an open question, because it is possible to improve each other, apparently, in different ways.

Then I came across kozul inference, where they do it in a similar way.

I don't know.

And that there is any use of such filtering by the second model?

1...310931103111311231133114311531163117311831193120312131223123...3574
New comment