Machine learning in trading: theory, models, practice and algo-trading - page 283

New comment
 
Vladimir Perervenko:


Indicators referred to by the common name "ZigZag" do not peek or move anywhere.

Sure, sure...

Good luck

 
SanSanych Fomenko:
I was interested in the meaning of the word "volatility. What exactly do you take as a measure of volatility?
RMS of returnees, sorry I was not accurate, not the value, but its normalized change, as well as with returnee (SDt - SDt-1)/SDt-1
 
Vladimir Perervenko:

As for which ZigZag values to use in training, there are three options :

  1. all
  2. all with increased example weights around the peak (if the model allows using a vector of example weights)
  3. only a few values around the peak
Depending on the model(s) you use, you can use one, and if the model allows for pre-learning, you can use two or all three in sequence.

Good luck

You forgot one more option.

4. ZigZag is not used in forecasting. :-) Even as a target function, it doesn't look good. The simplest and truest way, I assure you. For the forecast: Percentage of change over 10 bars forecast 1 bar forward. For the classification, the signal took a profit of 1 no 0. That's basic, so there's a bunch more target functions, but if you're predicting these two not a shitty one. It's about the entries, I'm telling you as a surgeon.

 
It's a rhetorical question:

Of course, of course...

========================================

This should be understood as: "I take back what I said about SanSanych's mistake. Wrong"? Or what?

The question is rhetorical.

 
Mihail Marchukajtes:

You forgot one more option.

4. ZigZag is not used in forecasting. :-) Even as a target function, it doesn't look good. The simplest and truest way, I assure you. For the forecast: Percentage of change over 10 bars forecast 1 bar forward. For the classification, the signal took a profit of 1 no 0. That's basic, so there's a bunch more target functions, but if you're predicting these two not a shitty one. Then it's about the inputs, I'm telling you as a surgeon.

Thank you doctor. Everyone has their own experience and approach.

Good luck

 

Vladimir Perervenko:

This should be understood as: "I take back what I said about SanSanych's mistake. Made a mistake"? Or what?

The question is rhetorical.


Exactly! You have opened my eyes to the truth! In "clever" book Safin repeated many times that the more complex the system, the higher the probability of overpotgonki, I am a fool, not only used a lot of indicators, but also neural networks, where thousands of parameters, and it turned ZigZag not potstvyat and can stupidly trade his knees!!!

Thank you!!! Just do not tell anyone this is a grail!

 
Vladimir Perervenko:

Thank you doctor. Everyone has their own experience and approach.

Good luck

Guys, I'm not bragging, do not think, but with networks closely work since 2006. Mostly first on the NS, then on the predikshina. We zigzag, and the grid from the grid and the indicator from the indicator. We've done just about everything. Believe me. There is a user Wizard, so this old fogey (no offense), because I remember him, too. There was an initiative group on a closed forum, gathered the real experts. It all started for me in 2007. So there just did not try. It has one very annoying feature - the current knee, which is re-drawn, and in fact you do not know when to start analyzing the market to meet the reversal. I am an adherent of course to classification, but I used to do forecasting as well. For this reason who predicts the market, try to predict the Percentage of change for 10 bars, at least one bar ahead. If the result is bad, let's think together how to organize good data. More precisely, I know what data may cause the price. That is what I want to try.

And also imagine that I have 10 inputs of 10 indicators and an output variable. All this is on a certain period of history. Using MT4 optimizer I adjusted parameters of indicators on this section to make it profitable. Then I applied the same indicators with adjusted parameters to the input of the Reshetov's optimizer. What do you think? Not only has the generalizing power not increased, but it has even worsened. Because to learn and generalize is not the same thing. So think about why it happened this way. It seems that indicators at this site are earning individually, but when the input NS was fed, then the generalization was not good. Why is it so, for me the question remains a mystery. So maybe someone here will be able to play the light. Thank you!

 
Mihail Marchukajtes:

Guys, I'm not bragging, do not think, but with networks closely work since 2006. Mostly first on the NS, then on the predickey. We zigzag, and the grid from the grid and the indicator from the indicator. We've done just about everything. Believe me. There is a user Wizard, so this old fogey (no offense), because I remember it, too. There was an initiative group on a closed forum, gathered the real experts. It all started for me in 2007. So there just did not try. It has one very annoying feature - the current knee, which is re-drawn, and in fact you do not know when to start analyzing the market to meet the reversal. I am an adherent of course to classification, but I used to do forecasting as well. For this reason who predicts the market, try to predict the Percentage of change for 10 bars, at least one bar ahead. If the result is bad, let's think together how to organize good data. More precisely, I know what data may cause the price. That is what I want to try.

And also imagine that I have 10 inputs of 10 indicators and an output variable. All this is on a certain period of history. Using MT4 optimizer I adjusted parameters of indicators on this section to make it profitable. Then I applied the same indicators with adjusted parameters to the input of the Reshetov's optimizer. What do you think? Not only has the generalizing power not increased, but it has even worsened. Because to learn and generalize is not the same thing. So think about why it happened this way. It seems that indicators at this site are earning individually, but when the input NS was fed, then the generalization was not good. Why is it so, for me the question remains a mystery. So maybe someone here will be able to play the light. Thank you!

Half a branch shed some light: predictors have no predictive power and are noise for the target variable. That's why the model is retrained, and the retrained model has NOTHING to do with its future use. NOISE IS NOISE ALL THE SAME, IN ONE APPLICATION THERE IS ONE RESULT, AND IN ANOTHER THERE IS ANOTHER.
 
toxic:
The RMS of returns, sorry I was not accurate, not the value, but its normalized change, just like with return (SDt - SDt-1)/SDt-1
If you develop your thought, you should take the coefficients given by GARCH, which is a very accurate characteristic of volatility.
 
SanSanych Fomenko:
If we develop your idea, we should take coefficients given by GARCH, which is a very precise characteristic of volatility.
You can, but as meta-features for the input, GARCH is linear (MOC) and based on very primitive features, i.e. not very smart, I have hundreds of times more features and the model is non-linear. Besides volatility itself is not used directly, I don't trade options because of their low liquidity on forex, it goes as an input to a higher level model.
1...276277278279280281282283284285286287288289290...3516
New comment
Reason: