Dependency statistics in quotes (information theory, correlation and other feature selection methods) - page 26

 
C-4:

The idea is good, here is a generated SB chart with the same volatility as EURUSD. Alexey, would you be so kind as to perform an analysis for it. Let's look at the differences if there are any.
I am calculating. What is the essence of the transformation? Did I understand correctly that this series is a random stray, but the volatility patterns are fully translated from the euro/dollar pair?
 

The result of this data is already ready.

As usual, histogram of mutual information values at lags of up to 500. (THE SIGN OF THE INCREMENTS HAS BEEN RETAINED.)

It looks like the real EURUSD H1 data. However, the sum of the mutual information on such a series is: 1.46 Bits - versus 3.57 Bits of the real series.

If the volatility has been transferred absolutely precisely, that's great! Because it turns out that besides volatility, the rest of the information (2.11 Bits) is related to the sign of price increase in the real series of quotes.

What do you think?

 

Guys, give me some indicator that you like and let me know what you are predicting with it, like a sign of increment or a price level. I will run this indicator with different parameters and calculate the mutual information. It would be desirable to send the file with quotes and calculation of formulas of indicator in Excel, it will be good for everybody.

And finally, my result can be compared to the strategy tester, if it is possible. This will be a step towards practice, won't it?

 
Candid:

Are you demanding strict stationarity from real processes? I hope not.

It's not really a question of doing it this way.

It makes no sense to build a model for a non-stationary process, but take and build one, but consistently simplifying the problem. Take some parts of a non-stationary process, e.g. trend, and see what is the residual. Then we take the next step. The goal is to get a stationary process for the residue and this is done only to ensure stability of the forecast.

If there is no goal of a stable forecast, you can run numbers back and forth.

 
faa1947:

Are you demanding strict stationarity from real processes? I hope not.

It's not really a question of doing it this way.

It makes no sense to build a model for a non-stationary process, but take and build one, but consistently simplifying the problem. Take some parts of a non-stationary process, e.g. trend, and see what is the residual. Then we take the next step. The goal is to get a stationary process for the residue and this is done only to ensure stability of the forecast.

If there is no goal of a stable forecast, you can run numbers back and forth.

It is also desirable to retain at least some informativeness in a broad sense in the resulting series, compared to the original one, otherwise the prediction will be difficult to interpret.
 
alexeymosc:
It is also desirable to retain at least some informativity in the broad sense in the resulting series, compared to the original one, otherwise the prediction will be difficult to interpret.

One of the requirements for such transformations is reversibility
 

Это лозунг, который говорит, что где-то что-то видел.

http://www.amazon.com/Introduction-Econophysics-Correlations-Complexity-Finance/dp/0521620082

Reading these publications doesn't make me any more money.

This is unfortunate.

 
alexeymosc:

The result of this data is already ready.

As usual, histogram of mutual information values at lags of up to 500. (THE SIGN OF THE INCREMENTS HAS BEEN RETAINED.)

It looks like the real EURUSD H1 data. However, the sum of the mutual information on such a series is: 1.46 Bits - versus 3.57 Bits of the real series.

If the volatility has been transferred absolutely precisely, that's great! Because it turns out that besides volatility, the rest of the information (2.11 Bits) is related to the sign of price increase in the real series of quotes.

What do you think?

The question is the sufficiency of the statistics. There is one realization of a random process with a given volatility profile. It is desirable to process several realisations and see the scatter.
 
alexeymosc:

If the volatility has been transferred exactly, that's great! After all, it turns out that apart from volatility, the rest of the information (2.11 Bits) is carried by the dependencies relating to the sign of price increments in the real series of quotes. What do you think?


The volatility is transferred absolutely precisely. In this chart the tick volume is completely the same as that of the hourly EURUSD bars, everything else is pure coin flip compressed into OHLC.

Alexey, is it possible to show the difference between the Rand EURUSD and the real EURUSD on the chart? I would like to see those 2.11bits first hand.

 
C-4:


The volatility has been transferred absolutely correctly. In this chart the tick volume is entirely the same as that of the EURUSD hourly bars, everything else is pure coin flip compressed into OHLC.

Alexey, is it possible to show the difference between Rand EURUSD and the real EURUSD on the chart? I would like to see those 2.11 bits for myself.

Good day!

I'm going to show you the comparison as well as a file with this data.

Blue series is calculated using your random data with volatility preserved.

This is a series of residues of mutual information of the real and random series:

Again, if the vola has been added exceptionally correctly, the visible residuals of the mutual information refer exclusively to the probabilities of the signs of the increments.