Machine learning in trading: theory, models, practice and algo-trading - page 3103
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
To the madness of the brave :)
This is where the development of all humanity stands.
If there is an understanding of the reasons why a model does not work, trained by standard methods, then we need to look for a solution to the problem. Let it not be perfect, but it allows to apply the model on the market with greater confidence and probability of success. Then already with the income to further study the problem and improve it.
If we consider neuronics, then perhaps it would be possible to divide the sample into bachs and increase the number of bachs (areas of change in the probability distribution) with the aim of their equal contribution to the learning process. With trees it is more difficult, although CatBoost uses similarly bachi on large samples, but it is not possible to control or manage. Although. there was a technology for continued learning - haven't experimented with it.... Have you?
Visually, the graphs are similar.)
Heteroscedasticity is modelled in econometrics and all kinds of applied statistics. There are a lot of tests invented there. R should have them all. The problem is that they give an estimate of the past and it is not certain that it is suitable for the current moment.
I don't see it the same way I see the use of SB.
For example, if I find a complex pattern in the market, I can generate an SB and see if it is there.
If it's not on the SB, it's kind of good, I've found a property that is inherent only to the market.
and if there is a pattern on the SB, it's bad? I don't know, is it bad that there is a pattern on both?
Well, I would like to read intelligent people who have already asked this question.
This is where the development of all humanity stands.
If there is an understanding of the reasons why the model does not work, trained by standard methods, it is necessary to look for a solution to the problem. Even if not ideal, but allowing to apply the model on the market with greater confidence and probability of success. Then with the income to engage in further study of the problem and improvement.
If we consider neuronics, then perhaps it would be possible to divide the sample into bachs and increase the number of bachs (areas of change in the probability distribution) with the aim of their equal contribution to the learning process. With trees it is more difficult, although CatBoost uses similarly bachi on large samples, but it is not possible to control or manage. Although. there was a technology for continued learning - haven't experimented with it.... Have you?
I've done variants with pre-learning, it didn't work that way. In bousting, the weights from past iterations don't change during chiselling, as in neurons, just on top of it. This is a disadvantage.
I've done neurons of all architectures too, including encoders-decoders for generating synthetic data. It's not much needed on the foreach either.I don't see it the same way I see the use of SB.
For example, if I have found some complex pattern in the market, I can generate SB and check if it is there.
If it's not there, then it's kind of good, I've found a property that is inherent only to the market.
and if it is on the SB, is it bad? I don't know, is it bad that the pattern is there and there?
Well, I would like to read intelligent people who have already asked this question.
Well, it's kind of a standard gentleman's trick for a trader-tester. You find a pattern in the market. Then you check it on a quote based on the SB. If the forecast on the SB is 50/50, then you can more or less trust the testing. If the percentage of forecast is about the same, then you look for where there is a look into the future. If there is none, then you look for a cunning peek into the future. If there isn't, you look for a very clever peek into the future. Something like that.
Well, it's kind of a standard gentleman's trick for a test trader.
To the madness of the brave :)
Discounted wreaths :).
Are we done with the PBO?
You've talked enough and forgotten?
Is the PBO done?
You've talked enough and forgotten?
Not forgotten, but discarded.
You should always test on two files.
The 1st one is divided into three parts by sample: 70%, 15%, 15%. On the first learn with cross-checking with min 5 folds and a sufficiently large fold. For RF it is 1500. Then we run on the second and third sample, and then on the second file, which is "as is". The classification errors on all samples should be about equal.
What will the RFO add to this?What will the RHE give in addition to this?
What it does is written in the article
All the pain of such A/B testing in one video
(don't watch for the particularly impressionable)