Machine learning in trading: theory, models, practice and algo-trading - page 769
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
Best R^2 score: 0.007975925
I guess this is the result you asked DR.Trader?
And now the most interesting part. I used this dataset to train the model and put it on the real. According to your approaches the quality was not very good, as I understand. But let's see whether the model is able to raise the score and if so, Reshetov's optimization method will be much better than all that you suggest by definition.
At the end of next week I will test the model created by DRTrader in P and we will see the error factor and also how the model created by Reshetov's optimizer works during this period. Then we'll see who is cooler...
I am quietly digging into Reshetov's code, he was really an ACC in programming. And realized that his method of finding a pattern in the data matrix was not the usual way everyone is used to look, going through the columns first, then rows, etc.
Reshetov's method of finding a pattern is complicated. I call it "square-nesting", who in the construction of printed circuit boards that would understand.
The trick is that the search is not linearly from column to column, but in a square way. I can't even articulate how it is..... maybe later I will try to describe it in detail...
Best R^2 score: 0.007975925
I guess this is the result that you asked DR.Trader?
Yes. This is quite a weak result, we may not even try to do forex with it. We should try to improve the set of predictors (remove/add), in order to make this estimate higher.
For example in the text file, the first column is the line numbers. It must be definitely removed.
Yes. This is quite a weak result, you can not even try to go to forex with it. You should try to improve the set of predictors (remove / add) to make this score higher.
For example in the text file, the first column is the line numbers. It definitely needs to be removed.
Well I removed it. By the way this training set which advised me vtreat, and Reshetovskiy optimizer when building models from it chose certain columns. What if you run only the columns that Reshetov selected. I'll try it tomorrow now I'm going to bed...
Best R^2 score: 0.007975925
I guess this is the result you asked DR.Trader?
And now the most interesting part. I used this dataset to train the model and put it on the real. According to your approaches the quality was not very good, as I understand. But let's see whether the model is able to raise the score and if so, Reshetov's optimization method will be much better than all that you suggest by definition.
At the end of next week I will test the model created by DRTrader in P and we will see the error factor and also how the model created by Reshetov's optimizer works during this period. Then we'll see who is cooler...
I am quietly digging into Reshetov's code, he was really an ACC in programming. And realized that his method of finding a pattern in the data matrix was not the usual way everyone is used to look, going through the columns first, then rows, etc.
Reshetov's method of finding a pattern is complicated. I call it "square-nesting", who in the construction of circuit boards that understand.
The trick is that the search is not linearly from column to column, but in a square way. I can't even articulate how to do it..... maybe later I'll try to describe it in detail...
If you noticed, he uses kernel tricks, i.e. even one feature can be used several times, but transformed in different ways
when you uploaded the source code of a neuron to mql it was visible there, maybe that's the key to success and slowness of calculations
basically, it makes almost no difference what you feed it as input, it will still come up with all the signs :)
(What fun you have here?))) Yes with me that all has long been clear. So show me a couple of weeks in the branch
"forex forecasts follow" manual intraday trading. Demonstrate the class so to speak ...
This is the theme of the MO, I'm not interested in manual discussion now.
You'd better send me something good in mql, for example, codes for RL, because I'll understand it before the bathing season starts.
♪ 'cause I've been shakin' in your supervised pipe ♪
Do we really need Polovtsian dancing? Maybe something simpler?
There, in the thread, Yusuf wrote
https://www.mql5.com/ru/forum/228879/page2
I did it for him... of course it is nonsense to analyze something by other instruments
But as for the analysis of the dynamics of LR coefficients (autoregression) in the sliding window, are there any studies about it?
it's something like ACF
that is, if we feed not only the autoregressive but also its coefficients into the model
There, in the thread, Yusuf wrote
https://www.mql5.com/ru/forum/228879/page2
I did it for him... of course it is nonsense to analyze something by other instruments
But as for the analysis of the dynamics of LR coefficients (autoregression) in the sliding window, are there any studies about it?
it's something like ACF
i.e. if we feed not only the autoregression but also its coefficients into the model
Sultanov is a separate, unique phenomenon in this forum.
But to use as predictors the parameters of any models, I have had this idea for a long time. But the problem is still the same: the relationship between such a predictor and the target variable should not change, and if it does, it should change slowly. And if it doesn't, then again it is NOT stationary, which excludes predictability of the future, even of the next bar.
There, in the thread, Yusuf wrote
https://www.mql5.com/ru/forum/228879/page2
I did it for him... of course it is nonsense to analyze something by other instruments
But as for the analysis of the dynamics of LR coefficients (autoregression) in the sliding window, are there any studies about it?
it's something like ACF
i.e. if we feed not only autoregression but also its coefficients into the model
I came to the conclusion that autoregression and autocorrelation function are different things.
Here is, for example, the acf of a trend area. The acf descends smoothly from the left edge and there is no repeated zero line crossing.
But here is the flat section. I can see the difference. I ran this design in the tester and did not get any improvement. It proves that the trend does not work in forex. So does a flat.
The trend often has no continuation and the flop is replaced by the trend. This is all the research, dotting the I's and crossing the T's...
I came to the conclusion that autoregression and autocorrelation function are different things.
Here is, for example, the acf of a trend section. The acf goes down smoothly from the left edge and there is no multiple zero line crossing.
But here is the flat section. I can see the difference. I ran this design in the tester and did not get any improvement. It proves that the trend does not work in forex. So does a flat.
The trend often has no continuation and the flop is replaced by the trend. That's all the studies that dot the I's and cross the T's...
the coefficient of autoregression of the 1st order coincides with the coefficient of autocorrelation of the 1st order, and they do not seem to coincide
well you can see in the picture on the left, only the charts themselves are different for some reason
ah, it's an akf both there and there, i got it
it's useless to use akf on nonstationary charts, you need to convert VR
I should ask SanSanych, he will explain you about acf for life and for acf ))