Machine learning in trading: theory, models, practice and algo-trading - page 1608

 
beat off through the signals in the cart? )
 
Maxim Dmitrievsky:
kicking back through the signals in the cart? )

as one option yes, through a couple of iterations

 
Evgeny Dyuka

- You don't need to look for sophisticated solutions, everything is simple, I got my first result with one-layer Sequential

- the forecast is composed of the aggregate opinion of 20 models


I think one of two is redundant here...

 
...:

- no need to look for complicated, contrived solutions, everything is simple, I got the first result on a single-layer Sequential

- the forecast adds up from the combined opinion of 20 models


I think one of the two is redundant here...

There is no redundancy. A model is such a small file - the result of network training. If you have a lot of them, you can manipulate them.
 
Evgeny Dyuka:

1. why an ensemble of models? what distinguishes them? is it the length of the future or the features?

or is there another network hierarchically higher, which chooses which network from the ensemble to listen to?


2. If all works well, then why the public? What prevents you from making money on your own?


3. completely candlesticks + indicators + something trickier and what is trickier? )

 
mytarmailS:

1. why an ensemble of models? what distinguishes them? is it the length of the future or the features?

or is there another network hierarchically higher, which chooses which network from the ensemble to listen to?


2. If all works well, then why the public? What prevents you from making money on your own?


3. completely candlesticks + indicators + something trickier and what is trickier? )

1. input data (chips)
2. As soon as you start thinking in terms of "earnings", everything immediately shrinks and stops, the brain switches to targets, stops, backtests, selection of parameters, while here we are interested in the theme itself, the research part.
3. this is know-how
 
mytarmailS:

1. why an ensemble of models? what distinguishes them? is it the length of the future or the features?

or is there another network hierarchically higher, which chooses which network from the ensemble to listen to?


2. If all works well, then why the public? What prevents you from making money on your own?


3. completely candlesticks + indicators + something trickier and what is trickier? )

2. I'd be happy to outsource the subject of trades on the exchange.
I can deliver neuro signals via sockets+json without any filters - as is, i.e. forecast every minute.
 
Evgeny Dyuka:
2. I'd be glad to outsource the subject of trades on the exchange.
I can deliver neuro signals via sockets + json without any filters - as is, i.e. forecast every minute.

You can translate trades here in MQL to "signals"

And immediately look for clients for subscription... I will be one of the first who will subscribe if I see that it is not a bullshit averaging machine.

 

I have a theoretical question

We have a target function to which we will approach the model

there are predictors, let them be 1000 pcs.


So the question is: if we have a lot of predictors can we divide them into equal parts, let it be 100 pieces and train 10 models.

Then outputs of these 10 models are fed to the new model as predictors. Will it be the equivalent of one model originally trained on 1000 predictors at once?

Something tells me that no, but I want to hear opinions

 
mytarmailS:

I have a theoretical question

We have a target function to which we will approach the model

there are predictors, let them be 1000 pcs.


So the question is: if we have a lot of predictors can we divide them into equal parts, let it be 100 pieces and train 10 models.

Then outputs of these 10 models are fed to the new model as predictors. Will it be the equivalent of one model originally trained on 1000 predictors at once?

Something tells me that no, but I want to hear opinions

it's called model stacking. Will not be the same, but not the fact that it will be more efficient. I did it this way, didn't see any improvement.

There is another way called meta-training. You train the first model to predict classes, then you get the results and feed them into the second model, on the same or different preictors, which allows/denies trading of the first model. 1 - trade, 0 - do not trade, depending on the quality of the predictions of the first model, i.e. a kind of filter. This greatly reduces the number of errors on the training data, but not so much on the new data (if the model has low generalizability). But in itself meta-training is a normal thing.

You can train the first model on some data, and the meta model on other data, on the errors of the first. There may be different variants. I did it both ways, in general there is an improvement, but it is more of a tweaking, rather than a way to get a quality model that will work on the OOS.

You can google Marcos Lopez De Prado "meta learning", just about trading