Machine learning in trading: theory, models, practice and algo-trading - page 460

 
Mihail Marchukajtes:

You're beating around the bush...... You tell me specifically, give me an example... Then we can continue the conversation, but otherwise....

He has no arguments, if he found something that he considers a mistake, and said about it Reshetov. And he in turn did not recognize it as a mistake, so his arguments are weak or wrong. So he keeps silent.


Arguments in the studio!

 
There's no argument to back up what you're saying.
 
Vizard_:

Mishek, why don't you say something? Support your brother)))

And what are you quacking here, do you have anything to say? At least people are doing something interesting.

Write something good

 

This is some very old version, and I 100% agree with Wizard that it's better not to use it. The model retrains, and also when at the end it shows the accuracy estimate on outofsample data - it shows it with an error, greatly overestimating it. I even posted in this forum thread an example where model in info after training showed 90% accuracy on new data, and after adding this data to formula and after calculating the result on the formula at least in excel the predictions were total randomness and only 50% accuracy

I had a lot of problems with jPrediction and some other things, I was working on it, I had a lot of problems with jPrediction and some other things. The site is no longer, ask Michael for the latest version and sources.
Imho this model turned out okay, but given the slowness of the work - in R there are things much more productive.

 
Dr. Trader:

This is some very old version, and I 100% agree with Wizard that it's better not to use it. The model retrains, and also when at the end it shows the accuracy estimate on outofsample data - it shows it with an error, greatly overestimating it. I even posted in this thread of the forum an example where the model in the info after training showed accuracy of 90% on new data, and when I pasted this data into the formula and calculated the result using the formula at least in excel the predictions were complete randomness and only 50% accuracy

Yuri then brought it to my mind, added a committee of several models, sped up and called everything jPrediction, created a website for this model. The site is no longer, ask Michael for the latest version and sources.
Imho this model is ok, but considering the slowness there are much more productive things in R.

I have, I also think that retrains, just a link to the description
 
Vizard_:

Then why are you yelling? He's not stupid. Wizard always writes about the point, even when he's being ridiculous.)
Throw out the rattle, do not waste time on it. Tools and their possible combinations voiced yesterday.
Don't waste your time on Mishek, he writes one thing implies another, and the res on the oos is the third...


I'm whispering. ) did not find the other posts, deleted by the moderators or something
 

I'm fine... I was just away for a little while....

Really, what's described on google is an old version of it.... BUT!!!!!!!

That it is reasonable to prove that JPrediction is retrained and does not work correctly, let's make an experiment. After all, everything is learned by comparison. This is exactly what I wanted to do.

Suppose we have a dataset, let's train it and let this model work for some time, and then let's see the result...

I train dataset on JPrediction, you train the same dataset on your AI, choose an interval and see which model will work longer and better.....

That's what I meant when I asked you to train my dataset on your AIs.

And so.... I don't know what criteria everyone used to decide that the predictor is overtraining????? Where did you get that from, Wizard. Do you have concrete evidence that the Optimizer isn't working???? I do???? Give me an example.....

Likewise, I can train your dataset and then you can see for yourself which model works better. The one you trained or the one I trained with the optimizer....

 
Vizard_:

No. Reshetov did not understand that we should not have fixed the normalization by the known formula. We should have made
a cover to disable it. Random breakdown is also questionable and I should have at least plugged the flag, but it's better off, etc...

Oh, yeah, I wrote about the breakdown, too. So for normal data ok, but specifically for forex needed to do some kind of roll-forward. Or at least just divide it into two parts by time - to train before the date, and to test after the date.

What's wrong with normalization? For neuronics it makes no difference what range the input is in, correctly initialized weights will digest everything. Normalization does not interfere, but also does not help.
Although intuition says that if the input contains positive and negative numbers, it is better not to shift zero. And R says to scale the predictors not to 0-1, but that sd(x) = 1

 
Vizard_:

No. Reshetov did not understand that we should not have made the normalization rigidly sewn up by the known formula. We should have made
a disconnect cap. Random breakdown is also questionable and should have at least flagged, but better off, etc...


About the random breakdown, I would argue.

When we make a prediction with AI, then YES, the sequence of data matters from the past to the future, because we are making a prediction.

But when it comes to classification, the sequence of data plays absolutely no role, because we have an area to divide and find the hyperplane that will do it best in the hope that the found law will be valid for some more time.

Just like the law we found when we built the prediction model.......

 

Not to mention the fact that the description shows that the model purposefully teaches those examples that are the hardest to learn, discarding those that are easy to learn. Well, that's just me... read from the description.... if I understand it correctly...

There, when splitting the sample into a training and test sample, it's like the two closest values fall into different samples. I understand that if we have two identical vectors, they will go into different samples, one in the training, the other in the test sample... So it's like this...