Machine learning in trading: theory, models, practice and algo-trading - page 696

 
elibrarius:

That's more accurate.

double pr2 = (pr!=0?log(pr):0)

Right, zero, thanks.

I want to try these things for classification tasks, they look nice

 
Vizard_:

Figure 1. Step 2, Step 3.))

I'm not as interested in the results as I am in the methodology of testing and comparing models. You should always remember that you should teach and test models on artificial input data, and on real data - it's a check with the result "this is how it turned out".

 
SanSanych Fomenko:

I am not as interested in the results as I am in the methodology of testing and comparing models. You should always remember that you should teach and test models on artificial input data, while testing them on real data is a "so-so" test.

What kind of artificial data is that? What do you mean by this data?

Surprised.

 
Guys and guys! My question is this. I use an indicator, I want to remake it into an EA, I found a video on how to remake it. The problem is that the indicator does not show up in the method I can not figure out what's wrong. I can't figure it out. Give me some direction on what to do.
 
Vladimir Perervenko:

What kind of artificial data is that? What do you mean by this data?

Surprised.

It's in the article.

The input data must have quite certain characteristics, and only then you can see the meaningful behavior of the model on:

  • sideways,
  • up and down trends
  • on breaks between sideways and trends and between trends
  • price spikes
  • on adjustable noisiness.


And then on what God has sent.

 
222ZXZ333:
Guys and guys! My question is this. I use indicator and I want to remake it in Expert Advisor, I found a video how to remake. The problem is that the indicator does not show up in the method I can not figure out what's wrong. I can't figure it out. Give me some direction on what to do.

In MT, you press the direct button on the indicator and select "change", if there is no such item, then there is no source and change it will not work.

 
SanSanych Fomenko:

This is in the article.

The input data should have quite certain characteristics, and only then you can see the meaningful behavior of the model on:

  • sideways,
  • trends up and down
  • on breaks between sideways and trends and between trends
  • price spikes
  • on adjustable noisiness.


And then on what God has sent.

This academic article examines artificially created time series with different, predetermined distortions with no reference to real data (a spherical horse in a vacuum). The aim of the study was to determine how different models react to such distortions in the TS, and which preprocessing operations affect the quality of the models.

The conclusions in the paper are:

  • Outliers significantly reduce the quality of model performance. They must be identified and handled. We know that
  • the use of first differences of predictors is useful. We also know that
  • Application of two consecutive first differences gives a significant positive effect. This is new. We need to check it out.
  • Using more than two consecutive first differences has a significant positive effect only for RF.
  • Theinput data needs to be normalized. This is also not new.

We have a lot of different input data from Forex market and stock market. Why do you need any artificial data? Take real data, transform it, reduce the dimension, etc., and use it to test various models. The rest is of the evil one.

Good luck

 

I just happened to learn how to predict the SMA(12). Here is a graph of the divergence between the moving average and its one-step-ahead prediction.

If you use such a predictor that looks one step ahead, then... It's not the price, though...

Or is there any other way to use it?

 
SanSanych Fomenko:

I just happened to learn how to predict the SMA(12). Here is a graph of the divergence between the moving average and its one-step-ahead prediction.

If you use such a predictor that looks one step ahead, then... It's not the price, though...

Or is there any other way to use it?

A few pages ago I wrote that predicting one step ahead is a no-brainer. About 70% of predictions are correct. The problem is that it's impossible to trade on that data. The thing is that the predictions are justified where everything is clear without them. That is, on smooth segments.

Tell me what TF you did 1 step ahead, and probably tomorrow or the day after, I will give a similar result to yours on the MA values prediction. I'm busy today.

 
Yuriy Asaulenko:

A few pages ago I wrote that predicting the MA step was a no-brainer. About 70% of the predictions are correct. The problem is that it is impossible to trade on this data. The thing is that the predictions are justified in places where everything is clear without them (predictions). That is, on smooth segments.

Tell me what TF you did 1 step ahead, and probably tomorrow or the day after, I will give a similar result to yours on the MA values prediction. Busy today.

Everything I try to predict usually has an error of at least 30%, and here is such an extraordinary thing, but useless

Well, maybe somebody will have some thoughts, future values are very important in machine learning.