There are a number of ideas as to why this might be useful:
1. Interdependent predictors can be identified:
1.1. Build a separate model with them and evaluate their predictive power
1.2. exclude them from the sample and evaluate their impact on the result, if they improve the result then consider creating similar predictors
Use one predictor instead of a group of predictors:
1. this will equalize the chances of taking it at random when building the model
2. Reduce training time by reducing dimensionality
Yes, I want to check this, but I don't know a tool for creating such a model out of the box.
There's a thought here by the way, why don't they use broken functions (like with quantization - instead of a step line) in training, it would allow for data precision lag and reduce overtraining.
If the predictors are prices, then 3-5 bars each on the left and right are very correlated with the center bar, exceptions are rare during sharp price spikes. Take every fifth to seventh bar and get roughly what you want. Or move to a higher timeframe. Or, filter out the predictors by checking their mutual correlation. The founder of this branch thinned the bars too, check his blog.
It's still the same... at the beginning of the week it works fine, after 'pre-training'. Then it starts pouring. I redid it again, tomorrow I will put it for tests :D
i have already tried it and it seems that the reward is incorrectly calculated in the trader... After a series of updates it starts trading in the wrong direction
I am also working on the recurrence nets in torch.
Maxim Dmitrievsky: Because it's impossible to train, the solver will get stuck in local minima. As for ideas - nothing can be taken out of there, because it is a black box
About jamming - perhaps need to change the way the error correction.
Why black box, if there are only 2-3 layers, it is quite real to unravel by coefficients. Small coefficients here can be coarsened and zeroed out, which will reduce the number of inputs to the neuron.
If the predictors are prices, then 3-5 bars left and right are very correlated with the center bar, exceptions are rare during price spikes. Take every fifth or seventh bar and get roughly what you want. Or move to a higher timeframe. Or, filter out the predictors by checking their mutual correlation. The founder of this thread also thinned bars, check out his blog.
Predictors are not prices in naked form - lots of relative points that may be similar...
I'm not sure that screening by correlation would be effective...
Allow the use of cookies to log in to the MQL5.com website.
Please enable the necessary setting in your browser, otherwise you will not be able to log in.
You are signed in. Two-factor authentication has been disabled. Set it up again and scan a new QR code, or continue with two-factor authentication disabled.
There are a number of ideas as to why this might be useful:
1. Interdependent predictors can be identified:
1.1. Build a separate model with them and evaluate their predictive power
1.2. exclude them from the sample and evaluate their impact on the result, if they improve the result then consider creating similar predictors
Use one predictor instead of a group of predictors:
1. this will equalize the chances of taking it at random when building the model
2. Reduce training time by reducing dimensionality
Yes, I want to check this, but I don't know a tool for creating such a model out of the box.
There's a thought here by the way, why don't they use broken functions (like with quantization - instead of a step line) in training, it would allow for data precision lag and reduce overtraining.
If the predictors are prices, then 3-5 bars each on the left and right are very correlated with the center bar, exceptions are rare during sharp price spikes. Take every fifth to seventh bar and get roughly what you want. Or move to a higher timeframe. Or, filter out the predictors by checking their mutual correlation. The founder of this branch thinned the bars too, check his blog.
I would recommend to start by clearing the data from trash attributes...
Imagine that you have 10 attributes, 9 of them are garbage attributes, you compress them all into one attribute.
It's still the same... at the beginning of the week it works fine, after 'pre-training'. Then it starts pouring. I redid it again, tomorrow I will put it for tests :D
i have already tried it and it seems that the reward is incorrectly calculated in the trader... After a series of updates it starts trading in the wrong direction
I am also working on the recurrence nets in torch.
yellow - beginning of weeks, first 1-3 days
I wonder...
Has anyone tried using "round levels" as signs?
Or as a way to handle prices ?
You can mark prices with circular values for example...
You can remove values that are the same in a row ...
It's not a bad compression of information, plus filtering ... Maybe it will be easier to look for patterns on such a chart for a model ...
Has anyone tried using "round levels" as signs?
Or as a way to handle prices ?
You can mark prices with circular values for example...
You can remove values that are the same in a row ...
It's not a bad compression of information, plus filtering ... Maybe it will be easier to look for patterns on such a chart for a model ...
I got a Renco
Has anyone tried using "round levels" as signs?
Or as a way to handle prices ?
You can mark prices with circular values for example...
You can remove values that are the same in a row ...
It's a good information compression, plus filtering ... Maybe it will be easier to look for patterns on such a chart for a model ...
Yes
hard redrawing
Because it's impossible to train, the solver will get stuck in local minima. As for ideas - nothing can be taken out of there, because it is a black box
About jamming - perhaps need to change the way the error correction.
Why black box, if there are only 2-3 layers, it is quite real to unravel by coefficients. Small coefficients here can be coarsened and zeroed out, which will reduce the number of inputs to the neuron.
If the predictors are prices, then 3-5 bars left and right are very correlated with the center bar, exceptions are rare during price spikes. Take every fifth or seventh bar and get roughly what you want. Or move to a higher timeframe. Or, filter out the predictors by checking their mutual correlation. The founder of this thread also thinned bars, check out his blog.
Predictors are not prices in naked form - lots of relative points that may be similar...
I'm not sure that screening by correlation would be effective...
I would start by recommending that you clean the data from the garbage signs...
Imagine - you have 10 attributes, 9 of them are garbage, you compressed it all into one attribute and what?
And what method for cleaning can you recommend?