Machine learning in trading: theory, models, practice and algo-trading - page 27
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
Hello!
I have one idea, I want to check it, but I do not know the tool for implementation... I need an algorithm that would be able to predict for a few points in advance, say for 3 or 5 (preferably a neural network)
Hello!
I have one idea, I want to check it, but I do not know the tool for implementation... I need an algorithm that would be able to predict for a few points ahead, say for 3 or 5 (preferably a neural network)
I've worked before only with classifications so I don't even understand how it should look like, advise someone how to do it or recommend a package in R
p.s. Great article Alexey
These are packages that extrapolate existing trends, such as forecast. The different splines are very interesting.
It looks pretty solid.
And what, there is no useful result?
When I ran the algorithm for the first time, on a small amount of initial data, there was no positive result, I got about 50% error with both y-aware pca and simple pca. Now I got a more complete set of data from mt5 - almost all standard indicators with all their buffers, some indicators are repeated several times with different parameters. I have created Expert Advisors for some indicators and used them to optimize indicator parameters for more profitable trading. On this data simple pca still makes 50% error, but with y-aware error in fronttest noticeably drops down to 40%. It's very interesting that the y-aware algorithm simply takes the raw data and makes a classifier that works correctly in 6 out of 10 cases. Conclusion - we need more raw data.
But this is where all the positive aspects end. You need 73 standard components for 95% accuracy. The predictor loadings in the components fluctuate from larger to smaller with no clear leaders. That is, there is no indication at all by which to select certain predictors. The model works somehow, but it is unclear what to do with it to improve the results or how to get the usefulness of predictors out of it.
The importance of the component:
loadings of predictors on the first 5 components:
ARIMA
But arima makes decisions by time series, and I need the model to make decisions from my data set, i.e. matrix with predicates and output a forecast for several bars ahead
No one will prevent you from training a neural network with several output neurons - each for a different planning horizon. At the same time it will be interesting to observe the results.
When I ran the algorithm for the first time, on a small amount of initial data, there was no positive result, I got an error of about 50% with both y-aware pca and simple pca. Now I got a more complete set of data from mt5 - almost all standard indicators with all their buffers, some indicators are repeated several times with different parameters. I have created Expert Advisors for some indicators and used them to optimize indicator parameters for more profitable trading. On such data simple pca still makes 50% error, but with y-aware error in fronttest noticeably drops down to 40%. It's very interesting that the y-aware algorithm simply takes the raw data and makes a classifier that works correctly in 6 out of 10 cases. Conclusion - we need more raw data.
But this is where all the positive aspects end. You need 73 standard components for 95% accuracy. The predictor loadings in the components fluctuate from larger to smaller with no clear leaders. That is, there is no indication at all by which to select certain predictors. The model works somehow, but it is unclear what to do with it to improve the results or how to get the usefulness of predictors out of it.
component importance:
Loadings of predictors on the first 5 components:
already did so, the neural network does not learn on a larger horizon with the target I set it
It's good that it did not learn, because you teach it on noise. But if it had learned, yes grail, yes on the real....
Busy here trying to eliminate noise. That's why we take so many predictors in the hope that at least something will remain.