Machine learning in trading: theory, models, practice and algo-trading - page 801

 
SanSanych Fomenko:

I calculated the predictive ability of 23 predictors for 12 currency pairs based on the difference in distributions for the teacher, with a correct prediction the profit would be more than 50 pips.

The results are as follows:

1. The predictive ability of the same predictors for different currency pairs is different.

2. Predictive ability of different predictors for one currency pair can differ by two orders of magnitude

3. Predictive ability will change as the window moves. As the window moves over 500 bars the statistics of variability of the prediction ability stabilizes

4. The slope of the predictive power obtained by moving the window varies from values of less than one percent to over 100 percent. And the "bad" predictors (with a large sko) are always bad, and the "good" predictors are always good.

5. We have studied 12 currency pairs. Three of them are hopeless: I couldn't find any good predictors for my target variable among 23 used ones.

6. For the same currency pair the predictive ability of longs and shorts is radically different.

What are they?

 
Maxim Dmitrievsky I was interested in your posts about reinforcement training. I have taken it to heart. I will now gather my strength and finally bring my genetic programming experiment to fruition. And then I'll see. The GP approach for trading has something in common with reinforcement learning in that there are no predefined target variables. Now we have to decide on the GP library. The main thing that it should have a normal documentation and make a strategy tester. I haven't found a good one. I will not lay out the code. I won't post the code and I won't share the ins and outs if there are any. This is not a good idea.
 
Oleg avtomat:

Which ones?

In principle it doesn't matter, since the pairs listed below are hopeless for MY predictors and MY target variable. These are H1: EURCAD, GBPJPY, USDCHF

It's quite possible that there are other sets of predictors on these pairs that will have acceptable predictive power for another target variable.

 
SanSanych Fomenko:

I calculated the predictive ability of 23 predictors for 12 currency pairs based on the difference in distributions for the teacher, with a correct prediction the profit would be more than 50 pips.

The results are as follows:

1. The predictive ability of the same predictors for different currency pairs is different.

2. Predictive ability of different predictors for one currency pair can differ by two orders of magnitude

3. Predictive ability will change as the window moves. As the window moves over 500 bars the statistics of variability of the prediction ability stabilizes

4. The slope of the predictive power obtained by moving the window varies from values of less than one percent to over 100 percent. And the "bad" predictors (with a large sko) are always bad, and the "good" predictors are always good.

5. We have studied 12 currency pairs. Three of them are hopeless: I couldn't find any good predictors for my target variable among 23 used ones.

6. For one and the same currency pair the predictive ability of longs and shorts is radically different.

Interesting, in p.3 we can assume that 500 is the lower limit of the training sample size, and in p.6 what difference do you mean, if by predictors, then it is not clear how it agrees with the fact that longs and shorts are in fact inversely proportional signals.
 
Grigoriy Chaunin:
Maxim Dmitrievsky I was interested in your posts about reinforcement learning. I took it to heart. Now I will finally gather my strength and bring my experiment with genetic programming to my senses. And then I'll see. The GP approach for trading has something in common with reinforcement learning in that there are no predefined target variables. Now we have to decide on the GP library. The main thing that it should have a normal documentation and make a strategy tester. I haven't found a good one. I will not lay out the code. I won't post the code and I won't share the ins and outs. This is not a good idea.

Yes, I saw your posts elsewhere (beginning) about GA, but then something so-and-so and forgot

So I understood that this is automatic programming with GA. The topic is actually close to RL, but the latter has its own advantages, e.g. guaranteed convergence to global optimum, unlike GA, + it's not quite clear how to bring in NS

To be honest I don't really understand how it differs from bot optimization with GA. I'll have to read it more carefully

 
SanSanych Fomenko:

In principle it does not matter, because further listed pairs are hopeless for MY predictors and MY target variable. These are H1: EURCAD, GBPJPY, USDCHF

It is quite possible that there are other sets of predictors on these pairs that would have acceptable predictive power for a different target variable.

I see. I see, thank you.

Sure, it's possible.

 
Maxim Dmitrievsky:

Yes, I saw your posts elsewhere (beginning) about GA, but then something so-and-so and forgot

So I understood that this is automatic programming with GA. The topic is actually close to RL, but the latter has its own advantages, e.g. guaranteed convergence to global optimum, unlike GA, + it's not quite clear how to bring in NS

To be honest I don't really understand how it differs from bot optimization with GA. I'll have to read it more carefully

There is no NS there at all. If it's simple, you use a genetic algorithm to derive the formula by which the TS works.
 
SanSanych Fomenko:

I calculated the predictive ability of 23 predictors for 12 currency pairs based on the difference in distributions for the teacher, with a correct prediction the profit would be more than 50 pips.

The results are as follows:

1. The predictive ability of the same predictors for different currency pairs is different.

2. Predictive ability of different predictors for one currency pair can differ by two orders of magnitude

3. Predictive ability will change as the window moves. As the window moves over 500 bars the statistics of variability of the prediction ability stabilizes

4. The slope of the predictive power obtained by moving the window varies from values of less than one percent to over 100 percent. And the "bad" predictors (with a large sko) are always bad, and the "good" predictors are always good.

5. We have studied 12 currency pairs. Three of them are hopeless: I couldn't find any good predictors for my target variable among 23 used ones.

6. Predictive power of longs and shorts is radically different for the same currency pair.

I can suppose from point 1 that the common factors acceptable for different instruments have been selected incorrectly for the predictors.

As for item 2 follows from the error in item 1.

On item 3, most likely the wave structure is not taken into account.

 
Ivan Negreshniy:

Interestingly, from point 3 we can assume that 500 is the lower limit of the training sample size

All figures given by me refer only to me - they cannot be generalized, although 500 for H1 is a little less than a week.

As for p.6, what kind of difference do you mean, if it is based on predictors, then it is not clear how it agrees with the fact that longs and shorts are in fact inversely proportional signals.

I do not understand it either, but it's a fact.

 

Hey guys!


Is the AI bot ready yet?

And how does it trade?

Just curious.


p.s. And give me your home address...