Machine learning in trading: theory, models, practice and algo-trading - page 1094
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
tell me - is it necessary to remove the trend from the price and if so, explain why
Oh, you're too much of a wordsmith...
If it were up to me, I'd permanently ban you so you wouldn't ruin the image of a decent forum
If you wait another few minutes you may start cleaning your posts as usual.
it's hilarious how accurately he can characterize you, not a single extra word, just the essence
You should be on the housewives' forums and not with the normal guys
I am not a teacher of explaining, I myself do not know))) Econometricians decompose, etc., and predict trend and noise, etc.
I showed the doc how you can forecast ox better than arima by means of mo, just make a bunch of features)))
From 1 VP you can make several thousands, some of them will give an advantage over ar... If you put
into rattles - it's desirable to standardize, though not into everything... You can generally go to another
feature space, etc... In short - whatever gives the best result is good. If you have an assumption
- just check it and that's it... and don't give a shit about listening or asking anybody. I have no trends, no transients
, no header at all... Just model timing. You can take from the code I posted
1 crap out of several allocated from VR, and try to forecast it or whatever you're doing.
Don't call it a rattle...
Well, econometricians would have removed it anyway, and if you do predictors, you should probably do it too, but here I have a slightly different case...
When I heard the word correlation several years ago I had a "fix idea" and it occurred to me - what if I look for patterns in history that are similar to the current situation and see how they had terminated in the past? The idea is interesting, and most importantly without parameters in the classical sense, the similarity decided to look for it through the Pearson's correlation, well it was green in general))))
By the way, it was a stimulus for me to learn how to program, I really believed in this idea ... Later it turned out that here on the forum, a few people also did so, and obtained the same results, then it turned out that there is a whole method that had long been invented by scientists called "LA" (local approximation), Ivakhnenko "MGUA" also has something similar "Method of complexing analogues ", in short, all been thought up before us)))
So the bottom line is that it doesn't work on the market, but it seems to me that I managed to make this method work, just all of the authors and I allowed one fundamental error, on Monday I'll verify the live market and then will share more structured information ...
If you're interested read "local approximation" method, it's in the end of this booklet. And tell me your opinion if it's necessary to remove the trend in this approach
read from page 103 20 Forecasting of natural origin time series
You should understand one simple thing: a trend is deleted so that there would be no bias in the model when another trend starts, and you were trained on another trend. If you can make it so that there is no bias, you don't have to delete it.
They will appear sooner or later anyway, because the markets change. The question is to find the balance so that something would work for some time.
A good cycle for the TS is 3 months, i.e. contract. Half training and half trading. This is from the results of training and tests. This is for futures on the forum. On other instruments it may be different.
For other symbols it may be different. Read the end of that booklet I threw ((local approximation)), I understand what you're talking about but in this particular approach maybe it doesn't work.
I just want to hear other opinions and argumentsread the end of that brochure I threw ((local approximation)), I understand what you're talking about but in this particular approach maybe it doesn't work
In the code base I found an indicator by Vladimir I think. Local approximation by the nearest neighbor method (or something similar), it also searches for similar regions in the history. Doesn't work, I think.
https://www.mql5.com/ru/code/133In the codebase there was an indicator lying around, the author of Vladimir, I think. Local approximation by the nearest neighbor method, it also searches for similar areas in the history. It does not work.
https://www.mql5.com/ru/code/133It does not work, it does not work, but it's too early to say....
The question is whether the trend should be deletedIt doesn't work, it doesn't work, but.... it's too early to tell
The question is whether the trend should be removedin this case yes, because Pearson's search on unsteady data is like s... against or downwind, and you need doldrums
in this case, yes, because Pearson's search on unsteady data is like s... against or downwind
Okay, there's an opinion. But there's a bunch of these distances, there's Kandel's rank correlation and a bunch more, there's Euclidean distance, there's 10 different subspecies there too, there's the dtw algorithm in the end
Okay, there's an opinion... But there's a bunch of these distances, there's rank correlation, there's a bunch more, there's Euclidean distance, there's also all sorts of subspecies of 10 things, there's in the end the same dtw algorithm
it's not an opinion but a given :) rank correlation will not save, what about the Euclidean distance dunno
This is not an opinion but a given :) rank correlation will not save, what does the Euclidean distance have to do with it?
Well the proximity / similarity, you can measure not just one correlation, there are a good 20 other ways
SZY
Ivakhnenko also writes that it is necessary to remove the trend, but he says that it is not necessary
quote: