Machine learning in trading: theory, models, practice and algo-trading - page 280

[toxic]

Vladimir Perervenko:

This does not prove anything.

The rule is simple: when preparing data for training , shift the target to the left ("into the future") by one bar, regardless of how you generated the signal. Do you have any idea why?

If you do not know, I will describe it in detail.

Good luck

Geez... you guys are so dumb...

You need to shift when you use signs of returns as a target, in a trivial case you have N of past returns as features, {Rt-n,...,Rt} and future Sign(Rt+1) as a target, then you shift to the left. But ZZ is ALREADY moving! HE'S PEEKING!

The peeking inductions don't need to be shifted, you then teach the classifier to be ahead of the future already, you can do that, but that's WAY worse.

That's all, I did not hire to convince anyone of anything.

Good luck.

[СанСаныч Фоменко]

toxic:

That's it, I didn't hire myself to convince anyone of anything.

Good luck.

We don't need convincing - we've got a lot of experience with ZZ.

The number of bars to shift the target to the left is the number of prediction steps. 10 steps shifted - 10 steps predicted.

Unfortunately, you don't understand that.

But there are much more important things than target shift.

This is especially true in the case of a CZ.

You will most likely not be able to use a phase in the way you have drawn it. No, it will teach you and you may even get an out-of-sample error of about 40%, but you won't be able to use it in real trading. The point is that in trend trading you need to predict a reversal, not the leverage itself. It's pretty in the picture, but why would you want to predict a trend at its end? Therefore, the target should be around the ZZ break.

This is the ideal, of course.

But as always there is a lot of fly in the ointment.

Several people here in this thread have tried to pick predictors for such a target and they failed. That is why it is not at all clear what is more important: the target or the predictors.

[toxic]

SanSanych Fomenko:

You don't have to convince us, we've been around ZZ for a long time. So, you'll have to be a little less stiff.

The number of bars to shift the target to the left is the number of prediction steps.

Unfortunately, you don't understand that.

I'm really far from you who ate the dog on ZZ, here I remember there were quants with funds, at least they can laugh, if anyone else is here, from the highlighted fragment, as well as from the "rule", that you have to shift no matter how filtered signal to make a target, even if filtering is a shift or more cunning peeping)))))))

[СанСаныч Фоменко]

I havenot yet succeeded:
I'm really far away from you who ate the dog on ZZ, here I remember there were quants with funds, at least they let them laugh, if anyone else is here, from the highlighted fragment, as well as from the "rule" that you have to shift no matter how filtered signal to make a target, even if filtering is a shift or more cunning peeping)))))))

For some reason you ignored the second part of the post.

Find the predictors for the target, which is located around the ZZ vertices. We have not yet succeeded.

[toxic]

SanSanych Fomenko:

For some reason you ignored the second part of the post.

Find the predictors for the target, which is located around the ZZ vertices. We have not yet succeeded.

I don't use ZZ knees as targets and I don't use the slope sign either, I don't useZZ at all. I joined this discussion to correct a minor error in the article, but I'm glad I was misunderstood, since we are not defending dissertations here)))

What drives the price? Informed participants, the big shots. When there is no new information from informed players, the market is in a flat. (information from KO) And now look at the chart, draw retrospectively, presumably, where the price started pushing the big guy and where he came out, and then look where the ZZ knees are and you'll understand everything...

[СанСаныч Фоменко]

toxic:

I do not use the knees ZZ as targeting and the dodge sign too, in generalZZ do not use. I joined this conversation to correct a minor error in the article, but I'm glad I was misunderstood, after all, we are not defending dissertations here)))

What drives the price? Informed participants, the big shots. When there is no new information from informed players, the market is in a flat. (information from KO) And now look at the chart, draw retrospectively, presumably, where the price started pushing the big guy and where he came out, and then look where the ZZ knees are and you'll understand everything...

Everything you write is clear and tempting, except one thing: how do you generate a target variable? On the history by hand? How long will such a target variable live in the future? If automatically, what is the algorithm for periodically updating the target variable to retrain the model?

I don't even see any ideas about this...

[Mihail Marchukajtes]

SanSanych Fomenko:

Everything you write is clear and tempting, except one thing: how do you generate the target variable? On the history manually? How long will such a target variable live in the future? If automatically, what is the algorithm for periodically updating the target variable to retrain the model?

I don't even see any ideas about this...

It's always a philosophical choice, choosing an output variable, and most importantly choosing one with a sufficient frequency level, when programming. You have to specify exactly what we want. First of all it is a profit. Signals with a profit should be marked with 1 and others with 0. If we are talking about the target function for prediction, then there is nothing better and easier than the percentage of change for 10 bars. It is enough to learn how to predict it for 1 bar (prediction window) and you'll be happy. If you can't do that, then the problem is in the input data.... Something like this.

[СанСаныч Фоменко]

Mihail Marchukajtes:
It's always a philosophical choice, the choice of output variable, and the most important is to choose it with sufficient frequency when programming. You have to specify exactly what we want. First of all it is a profit. Signals with a profit should be marked with 1 and others with 0. If we are talking about the target function for prediction, then there is nothing better and easier than the percentage of change for 10 bars. It is enough to learn how to predict it for 1 bar (prediction window) and you'll be happy. If you can't do that, then the problem is in the input data.... Something like this.

We take ZZ - clearly and very accurately in the sense of reversals.

For example, we put 25 pips in the ZZ parameters, which guarantees ZZ reversals of more than 25 pips.

Then we form the target variable as follows.

For a given PZ reversal, we take the preceding PZ reversal and surround it with all bars whose price differs by more than 25 pips. Around this previous PZ reversal, we accumulate a certain number of bars (different for different reversals) which predict a future rise of 25 pips to the reversal in question.

Is it so?

[mytarmailS]

I decided to play with the package twitteR, it can be used to mine a text from tweeter, and using it you can try to identify some mood of the global community and so on, that is the idea to do something like Oanda, but in the fundamental analysis and not in the technical analysis.

But here it is like in the story - a guy was on his way to success but he stumbled :)

I got hung up ((, at the initial stage I can not connect the package. Karoch who is less in this topic and knows English and hum look at this pagehttps://www.credera.com/blog/business-intelligence/twitter-analytics-using-r-part-1-extract-tweets/ end of the third paragraph, the settings in the browser generates a pin code where the hell it should enter? already sick of banging my head against the wall, tell me who understands.

[Mihail Marchukajtes]

SanSanych Fomenko:

We take ZZ - clearly and very accurately in the sense of reversals.

For example, we put 25 pips in the ZZ parameters, which guarantees ZZ reversals of more than 25 pips.

Then we form the target variable as follows.

For a given PZ reversal, we take the preceding PZ reversal and surround it with all bars whose price differs by more than 25 pips. Around this previous PZ reversal, we accumulate a certain number of bars (different for different reversals) which predict a future rise of 25 pips to the reversal in question.

Is that how it works?

I confess I didn't quite understand what you wanted to say. Or rather did not understand it at all :-) In principle, you can use ZigZag as an output variable, but in a very narrow context. The only thing you got right, is that you need to choose the moment for analysis. At what point will you analyze the reversal of ZZ???? If every bar, then I'll tell you right away this is utopia. The market should be analyzed at the moment of its refraction. When the refraction has already occurred, because there will always be a place for a pullback, from which we will enter. Is not it?