Machine learning in trading: theory, models, practice and algo-trading - page 1487
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
That's why we need to give an estimate at the moment of sorting predictors and their values and not take those predictors with values (ranges) that are very crowded, but give preference to those that are distributed over the whole sample.
I assume you need a uniform forward? The forest doesn't know anything about forward in training\.
Well, everything always looks very good on the backtest)
I guess you need a uniform forward? The forest doesn't know anything about the forward in training\.
The assessment must be on the two samples participating in the study.
The evaluation should be on the two samples involved in the training.
Then there will be a fit to the validation plot. Although it may work if validation is larger than training. In Ivan Butko's NS, training on a small plot, and validation on a plot several times larger. Here we usually considered options with validation of 15-25% of all data, and in his video about 80%
Is the second one validation?
Yes, if it applies.
In general, I think the larger the sample the better - it's a question of predictor stability, not just learning.
It might be possible to preprocess the data, to cluster just the ranges of predictor values that are not rare, but are heavily crowded across the sample.
Replace them conventionally with -1 and disallow doing initial splits on them, at least to a depth of 2-3.
Then there will be a fit to the validation plot. Although it may work if validation is larger than training. In Ivan Butko's NS, training on a small plot, and validation on a plot several times larger. Here we usually considered options with validation of 15-25% of all data, and in his video about 80%
There won't be much of a fit, because we don't really train on validation, we just take additional information about the market structure from it as well.
Yes, if it applies.
In general, I think the larger the sample the better - it's a question of predictor stability, not just learning.
It might be possible to preprocess the data, to cluster just the ranges of predictor values that are not rare, but are heavily crowded across the sample.
Replace them conventionally with -1 and forbid making initial splits on them, at least to a depth of 2-3.
There will be no special adjustment, because we are not learning on validation, but simply take additional information on the market structure.
And we come to the point that we just need to include validation in the training section.
No, what kind of svm is this "Parzen window", kernel smoothing, and"quasi-optimal" in the sense that almost perfect (Mitchell had it somewhere), but only very slow, each iteration - sorting the whole dataset to a new point and convolution with the kernel
I'm not sure what effect it will have on non-stationarity... and if it's slow, it's hard to even montecarrelate
imho the main problem is scaling/transforming data, loop extraction since even non-linear regression or SVM give good results if patterns are repeated (on artificial BP)
i.e. the problems with pattern selection are farfetchedthe main problem is scaling/converting the data, highlighting loops
Yes, that's right.
Too complicated... The only acceptable variant is if validation is commensurate with or greater than training.
And we come to the point that we just need to include validation in the training section.
Standard algorithms are designed to work with stationary phenomena, closed systems, so there any information is considered a priori useful and does not evaluate it in terms of randomness, but only the possibility of use for the task (classification by target), while we have a lot of noise and I proposed a logical way to combat it.
And yes, I moved away from validation per se when collecting the herbarium, leaving the criteria for evaluating leaves, including efficiency on temporal sites. Now I have more than 50 thousand leaves from all tree variants, of which only about 200 are selected as signal and 600 as filters, significantly improving results (on most periods) (3 per signal leaf).
We have to assume that all we can do is to develop an algorithm for the most successful data fitting, because we do not know the future and there are a lot of variations, even based on the available predictor values. And, if we are lucky, we will be able to identify a pattern that will continue to exist for some time, so it is important to look for such a pattern with certain criteria, and logic suggests that at least it should be a pattern occurring throughout the entire sample. That's why the task is to create an algorithm that creates many such patterns and builds trees from them.
A simple example of perootooting from which we can protect ourselves is detecting a part of a sample with a certain range of prices at which a long flat has been fixed - when training now we will get two price ranges (or returnees from the upper TFs) to trade at, but it is unlikely that a flat will be fixed at the same place in the future. And with the approach suggested by me, with the same features, more likely will be found strike levels, as meaningful for event generation.