Machine learning in trading: theory, models, practice and algo-trading - page 728

 

And to summarize. An interesting fact was observed. Both models were trained on the same area of 40 records. However model with high value of VF, showed worse results than model with low VF, I connect it to the fact that for the second model, where the VF was small, and the results of learning high for this very model had to increase the interval of training, so to speak to give load to the model, because just in this area input data too well described the output and the model obtained TOO good. Now for the philosophy.....


Imagine a scale along the axis of the ICS. This scale is unique for each data set and somewhere on this scale there is a boundary, a vertical line, where the overtrained zone is on the right and the undertrained zone is on the left. The task of any optimization algorithm is to come as close as possible to this boundary from the side of undertraining, but not to jump this boundary. And the closer the algorithm gets to this zone, the less undertrained it becomes, while being to the left of the overtrained boundary. I understand it is difficult to represent in text form, but try to...... I actually have a theory about this topic, with zones. In general it is not so simple, but not the essence.....

If you don't look at this boundary strictly, then model training comes down to the optimal balance between undertraining and overtraining. That is, there has to be some balance. Returning to our model. It learned the input data well, because it matched the output, but it didn't let the model learn for the feedback, because it would have been enough to include a couple more paternals into training, which would have let the model perform even worse, but with added paternals that could have been decisive for the feedback.

In other words, if the model learned the data too well, it is necessary to increase the training period, thereby loading the model.

According to Reshetov's classification.

the first model 77-80%(VI 0.86) generalization, the second 88-90%(VI 0.65). Optimal variant of generalization level is 75-85%

 
Mihail Marchukajtes:

And to summarize. An interesting fact was observed. Both models were trained on the same area of 40 records. However, the model with high value of VI, showed worse learning results than the model with low VI, I

What is the VI? Maybe I can guess it off the top of my head. Time interval.

Yasha said:military publishing

 
Evgeny Belyaev:

What is VI? Maybe I can guess it off the top of my head. It's a time period.

Yasha gave me a hint:military publishing house.

Mutual Information.....

 

Once again, for those who tank: you're doing curvafitting on very short time intervals with a very small unrepresentative number of trades

That's not even for machine learning, but for the "Interesting and humorous" section :)

You keep carrying water in the sieve (excuse me, sieve :)), and then you are honestly surprised that there is no profit in the real account.

Well, make at least 1000 trades and then wonder why only the first 10 trades normally work at the CB occasionally, patch your sieve
 
Maxim Dmitrievsky:

Once again, for those who tank: you're doing curvafitting on very short time intervals with a very small unrepresentative number of trades

This isn't even for machine learning, but for the "Interesting and humorous" section :)

You keep carrying water in the sieve (excuse me, sieve :)), and then you are honestly surprised that there is no profit in the real account.

Well, make at least 1000 trades and then wonder why only the first 10 deals work fine on the OOS occasionally, fix your sieve

Let's wait and see.... My point is that a month of work at 15 minutes with more than 70 trades is not a short time interval.

Let's see how you sing when the result is transferred to the account.......


This just once again proves that giving a person a tool in his hands is not the fact that he will be able to use it correctly, treating it as a trinket......

 
Mihail Marchukajtes:

Let's wait and see.... I think a month of work on 15 minutes with more than 70 trades is not a short time interval.

Let's see how you will sing when the result is transferred to the account.......

I don't know why you're so sensitive to your own experience :) This program was written in 10 times shorter time than you're trying to take it to different places
 
Maxim Dmitrievsky:
God, why are you all so slow on the sensitivity to your own objective experience :) This program was written in 10 times less time than you try to apply it to different places

As you like, the main thing is that it does not retrain. In any case, it generalizes well enough, but I have nothing to compare it with, because I did not get to the networks in R.

I always proposed tests to compare your AI and Reshetov's optimizer model. But no one took the risk. Probably felt that you will lose....

 
Mihail Marchukajtes:

As you like, the main thing is that it does not retrain. In any case, it generalizes well enough, but I have nothing to compare it with, because I did not get to the networks in R.

I always proposed tests to compare your AI and Reshetov's optimizer model. But no one took the risk. Probably felt that you will lose....

Just tell me that you can not write a test for at least 1000 trades, 10 of which will give you a profit on the OOS on the real. And what you're doing is not even a backtest, okay? Increase the sample, or you'll be stomping around until the end of time.
What is the point of competing in something if you don't even understand the basics
 
Maxim Dmitrievsky:
Just tell me that you can't write a test for at least 1000 trades, 10 of which will give you a profit on the OOC on the real. What you're doing is not even a backtest, okay? Increase the sample, or you'll be stomping around until the end of time.
What's the point of competing in something if you don't even understand the basics?

Okay, Maximka, stop your hysterics here. Take a deep breath.... Exhale and exhale again... exhale..... and now watch the signal..... The coolest proof of all...

 
What hysteria. ) You are being told your case. A test on a long stretch of history, a forward test, and then we can talk about something. When training a network, it should be divided into training and test samples. I can say with almost 100% certainty that your signal will show a loss. 0.0000000001% for everything to work out.