Dependency statistics in quotes (information theory, correlation and other feature selection methods) - page 22

New comment
 
alexeymosc: Maybe, Alexey. At least the standard TFs can be tested in 3-4 hours. Is there even a way to unload custom timeframes from MT for analysis?

I don't know about custom ones (non-standard ones like M10, M20, etc.), but standard ones are always possible. On the M5 I can try to check it myself on my chi-square. It will take a long time to calculate, but art requires sacrifice.

It seems that custom history can be unloaded using period_converter script.

joo: I highly doubt it's M1. On M1 I haven't been able to get an acceptable trainability of the grids.

Andrew, your "target function" of the grid is probably not informative.

I'm not agitating you to train the grid on M1. It's just a matter of principle to check the dependencies on the smallest possible TF. Anyway, the percentage of bars on the history, on which zero depends, clearly increased in the sequence D1 -> H4 -> H1 -> M30. I have not checked at the smallest ones.

It would be too simple if all information in quotes was transmitted only through volatility. I don't believe that the market is so primitive that it is stupidly reduced to econometric models with heteroscedasticity.

 
joo:

Indicators are not used at all lately (if the term "indicator" is understood as a procedure for converting a quotient into a form which gives fewer signals than the number of bars in the reporting period).

2. No, but I'd love to try it, if I understand how to do it, having read this thread.

And if I want to somehow combine information by several bars, then the transformation will become a technical indicator?

2) I tried to do it. I performed a simple experiment with a mask. How do you usually do it? If the wave goes up, the price will go up. So I took the difference between adjacent MA values and looked for the maximum informative value of the price change sign. It turned out that it wasn't the last change, but tens of bars away. I also looked to see for which particular price change the MA is most informative, tried price changes one bar ahead, 2,3 ... 10. The maximum turned out to be 5 bars ahead, the MA was with a period of 5. It may be just a coincidence... But, it is important.

Although, of course, in MT it can all be run by mechanical overshoot.

 
Mathemat:

I do not believe that the market is so primitive that it can be reduced to econometric models with heteroscedasticity.

I have used 5 different tests for different types of heteroscedasticity for a large number of models and never found this heteroscedasticity.

 
Mathemat:
I suspect it's the M1.

On small timeframes you will need to get the prices right for the calculations. Doing them only by Bid is not good. Better 0.5*(Bid+Ask).
 
Mathemat:


I'm not agitating you to train the grid on M1. It's just a matter of principle to check the dependencies on the smallest possible TF. Anyway, the percentage of bars on the history, on which zero depends, has clearly increased in the sequence D1 -> H4 -> H1 -> M30. I have not checked at the smallest ones.

It would be too simple if all information in quotes was transmitted only through volatility. I do not believe that the market is so primitive that it is simply reduced to econometric models with heteroscedasticity.

Alexey, I agree! I'll do a measurement right now for hourly bars to compare to daily and 5 minute bars.

And I don't believe that all the different mutual information from noise has been reduced to volatility. Here's an example, and the basis of my doubts.

This is what the frequency matrix looks like (1st lag is the target variable) for random data with 5-minute characteristics.

We see that the probabilities are uniform, as expected for such data.

And this is what the matrix looks like for natural data:

We see that 1-5 and 5-1 are the stand out frequencies (although 5-5 also stands out because of the bunchiness of the volatility). And, while that's exactly what you won't make dough on, it's likely that the subject of interest is there. And if you take a few lag variables, you'll see even better prediction reliability. (By the way, note that the cross-entropy of the system for real data is smaller, i.e. predictability has increased, as it should. This relates to past discussions about a single number describing the whole system.)

Next, I'll take clocks and then take return values modulo, subtract one from the other and see what remains (there must be mutual information relating to the sign of price change). If I manage to do it today, I will do it today, if not, I will do it tomorrow.

 
faa1947: Used 5 different tests for different types of heteroscedasticity on a large number of models - never found this heteroscedasticity.
What was Ingle given the Nobel Prize for then( see 2003here )?
 
faa1947:

I do not believe that the market is so primitive that it is stupidly reduced to econometric models with heteroscedasticity.

I used 5 different heteroscedasticity tests for a large number of models and never found heteroscedasticity.


True, the market is more complex. However, this is no reason to ignore the observed phenomenon.

About the tests: heteroscedasticity is a widely accepted fact in the literature, which can be seen even by eye. If you cannot find it, it means that you did something wrong. Sometimes heteroscedasticity tests are applied to predictors and model errors, but this is more a check of model specification.

 
Mathemat:

Andrew, your "target function" of the grid is probably not informative.

I'm not agitating you to train the grid on M1. It's just a matter of principle to check the dependencies on the smallest possible TF. Anyway, the percentage of bars on the history, on which zero depends, has clearly increased in the sequence D1 -> H4 -> H1 -> M30. I have not checked it on the smallest ones.

It would be too simple if all information in quotes was transmitted only through volatility. I don't believe the market is so primitive that it would be reduced to econometric models with heteroscedasticity.

No, I actually have "nothing against M1". However, all other conditions being equal (at the same time I have observed the best results on H1, when the data is input in such a way that there is no information about volatility). So I said, maybe there is another TF, different from H1, somewhere nearby, which is "better".
 
alexeymosc:

2. and I tried to do it. I did a simple experiment with a waving machine. How do you usually do it? If it goes up, the price will go up. So I took the difference between adjacent MA values and looked for the maximum informative value of the price change sign. It turned out that it wasn't the last change, but tens of bars away. I also looked to see for which particular price change the MA is most informative, tried price changes one bar ahead, 2,3 ... 10. The maximum turned out to be 5 bars ahead, the MA was with a period of 5. It may be just a coincidence... But, it is important.

Although, of course, in MT it can all be run by mechanical overshooting.

That's why I'm interested in your research, because it answers the question: "What are the best bar numbers to feed to the grid and in what combination?
 
Mathemat:
And what was Ingle given the Nobel Prize for then( see 2003here )?
I mentioned five different tests - seems to be needed for something too. There's a funnier fact: the Matlab toolbox called "econometrics" only looks at different ARCH models. I've never traded options. Maybe there. But on forex and watched some stock instruments where the level of an instrument is modelled - not once. And when you create a model, you consider a very large number of options before you get something worthwhile - and not once. Although maybe I'm not good at it or maybe I'm in the wrong class of models. By the way, there was once an article here about ARCH modelling, so there was also a comment there that it wasn't applicable to us.
1...151617181920212223242526272829...74
New comment