Machine learning in trading: theory, models, practice and algo-trading - page 388

 
Maxim Dmitrievsky:


I already have predictors, strange as it may seem. I have a ready-made bot, which I built in less than a month. The most important thing - the predictors, it's out of the question, yes. For example, with my feverish fantasy predictors are chosen at once, I've been working as an analyst for 5 years :) In my opinion, selection of predictors is not such a difficult task as studying NS architectures, the main thing is to sit down and pick, take 2-3 weeks :)



And in numbers?

On the training sample, the test sample, and the validation sample.

Most importantly: on a new file that was originally separate from the previous three.

All these four values should not be very different from each other. If the error they differ by more than 10% (deviation of 30% and 35%, for example), then off to hell.


And what is worth real money is nothing, signals die after a year or even after two years.

 
SanSanych Fomenko:

What about the fact that increments do not indicate trends in any way?

Yes, they don't.

Either the model predicts the increment or the direction - that's what classification models are for.

I'm not aware of any classification models that recognize movement on the news. And for GARCH, that's the point of the model - to work out the occurred movement. Fat tails - this is the movement on the news when trends break and sharp reversals occur.


Well, you can watch the increase on different timeframes.

There are interesting GARCH models for several timeframes. The meaning is the following.

Suppose we predict the increment on H1. The model requires input data characterizing the distribution. As such input data (usually volatility) we take not the previous hour but the minutes within the current hour.

In my opinion, it is important to divide the entire history into sections with similar behavior. For example, here is a picture of EURUSD closing price for 5 years. It shows that there was one trend before Q2 14, then the rest of 14 and beginning of 15 is different and after that the third one started and is still going on. Mixing everything in one pile, like trying to get an average hospital temperature and use it to diagnose the condition of an individual patient - imho, wrong.


If we take, for example, the current trend, somewhere from early 15 to the present day, and at least just isolate/extrapolate trends, periodicity, you get imho, quite a plausible result. Here's a picture, in green is the closing price forecast for the next couple of weeks.


 
SanSanych Fomenko:


And in numbers?

On the training sample, the test sample, and the validation sample.

Most importantly: on a new file that was originally separate from the previous three.

All these four values should not be very different from each other. If the error they differ by more than 10% (deviation of 30% and 35%, for example), then off to hell.


What is worth real money is nothing, signals die after a year or even after two years.


I don't need so many useless samples, the training and testing ones are enough, I use GA to choose the parameters, then I choose the results, which are maximum similar to back and forward. You will never train a model for the entire history of quotes, plus you offer as much as 3 independent periods, and trade on the 4th, this is nonsense in the case of trading in the market, because the market changes during this time. So, just make sure the model is not overfitted, on the section outside the training sample, and that's it.

I retrain every week, so far the second week is +35%. What's on the real is about what, this is real money)

 
SanSanych Fomenko:

And what stands on the real - is nothing at all, there are signals, they die after a year or even two ...

Do you seriously want to create a market model for years to come...?
 
Maxim Dmitrievsky:


I already have predictors, oddly enough. There is already a ready-made bot, which stands on the real, I wrote in less than a month. The most important thing - predictors, it is not discussed, yes. Well, this is someone who has experience... For example, with my feverish fantasy predictors are chosen at once, I've been working as an analyst for 5 years :) In my opinion, selection of predictors is not such a difficult task as studying NS architectures, the main thing is to sit down and pick, take 2-3 weeks :)

Tell me please, what predictors do you use?
 
Maxim Dmitrievsky:


In the numbers allenorm, do not need so many useless samples, training and test this is enough, through GA selected parameters, then I choose the results, as similar to the back and forward. You will never train a model for the entire history of quotes, plus you offer as much as 3 independent periods, and trade on the 4th, this is nonsense in the case of trading in the market, because the market changes during this time. So, just make sure the model is not overfitted, on the section outside the training sample, and that's it.

I retrain every week, so far the second week is +35%. What I mean by real trading, this is real money.)

You know best about the samples.
 
pantural:
Tell me please, what predictors do you use?
I have already described one here, it is the value of the slope of the regression line and I even threw down an example of a bot, the others are a secret :)
 
Ivan Negreshniy:
Do you seriously want to create a market model for years to come...?

No, of course not.

I'm busy getting some guarantees for some future.

 
Maxim Dmitrievsky:


In the numbers allenorm, do not need so many useless samples, training and test this is enough, through GA selected parameters, then I choose the results, as similar to the back and forward. You will never train a model for the entire history of quotes, plus you offer as much as 3 independent periods, and trade on the 4th, this is nonsense in the case of trading in the market, because the market changes during this time. So, just make sure the model is not overfitted, on the section outside the training sample, and that's it.

I retrain every week, so far the second week is +35%. What's on the real is about, it's real money )

I also have two plots.

The first plot: three samples are randomly taken from it and taught, tested, and tested. The last section, which follows the first, is a sequential run, preferably with a tester.

I completely forgot, although I have written many times before.

The step described above is the second step.

The first step is the selection of predictors "relevant" to the target variable. I can prove that very good results are obtained with predictor sets where predictors that have nothing to do with the target variable - noise - prevail. Very good results are obtained on noise during training. Moreover, I managed to get an error of less than 10%, up to 3%, on the first part of the three parts mentioned above! And then I got a completely random error on the second part.

If you start to weed out the noise predictors, the error increases during training, but decreases in the second section. If you get rid of the noise predictors, you get about the same error value. On my set of predictors it is a little less than 30%.

 
It is not necessary to train machines, first of all, you need to have iron nerves and connections in the higher echelons of power, that would be profitable to trade