Dependency statistics in quotes (information theory, correlation and other feature selection methods) - page 72
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Exactly, that's exactly right. The ACF is useless.
The mutual information, however, should be, as there is no smell of zero even at a distance of hundreds of bars.
please send me your results - how you calculated it, what data you used, etc.
I need you to send me the results - how you calculated it, what data you used, etc.
I haven't counted the take-aways. I have other things - statistics. The chi-square criterion of the relationship between series of data. I'll post it later. I'll have to clarify it to make it clear.
In fact, this is very close to what the namesake does. Even the formulas are very similar there.
The maximum may be 2.098 bits. This is the average information of this particular data series. If, for example, a bar on lag 1 completely determines a zero bar, then their mutual information will become 2.098 bits.
What is this number? It is a measure of information ) You need to read articles on TI. In short, bits reflect a measure of the randomness of the data source values according to the eigeninformation formula of one particular value
I(X) = - log(P(x))*P(x).
Another example. We flip a coin, count the mutual information between two consecutive events. By formulas I've translated in my article we get that the mutual information I(X;Y) = 0. And if a tails roll precisely indicates a subsequent tails (or heads) roll, then I(X;Y) would be equal to 1 - this is the average information of the "fair coin" data source.
Alexey! I too use mutual information to select meaningful inputs for a neural network. I usually get models with predictability of 52 - 54%. It seems to me that worthwhile results will be obtained when mutual information is > 0.1 bit. For volatility it is possible to find such significant inputs.
I realise that my posts in this thread are on the verge of being foul, almost off-topic.
On the five, hrenfx.
The posts are fine, quite on topic. Also just about the applicability of TI to the subject, here some have doubts.
- That there is any invariance in the change of scale, sorry, I don't understand. I understand invariance as the presence of a scaling factor (in the general case it can be any number or function), when multiplied by which we get a new pattern of a different scale. That is, the affine transformation, which is the manifestation of structuredness in a chaotic flow of data. The problem then comes down to finding such a coefficient. When a pattern is found it is simply multiplied by this coefficient. And such transformation works both "upwards" and "downwards". And that's all.
Ok, let it be so, I overdid it with fractality. It is there, but it's not perfect. More precisely, fractal invariance is not perfect.
- if you investigate the relationship between the two quantities.
- why this is so, what causes this statement
What statement - I do not understand.
- Why two, and not three-five-thirty
- Which two values?
Here it is clearer.
1. Two because we take a conditional source and a conditional receiver and try to find out if there is any dependence between them.
We make up an alphabet (I divide the distribution of returns into quantiles, it's more convenient for me; the namesake does it a bit differently, but it doesn't affect the result much), apply TI. There are some doubts about the communication channel. Perhaps time is the communication channel.
If we fix the difference between bars (numbers are calculated in MT4) equal to, say, 238, then the source is the series
return(Bars-1), return(Bars-2), ... return(238) (about 80 000 values at 12 years of watch)
The receiver is a series of
return(Bars-1 - 238), return(Bars-2 - 238), ... return(0).
In short, just a series of returns and the same, shifted relative to itself by 238.
You can calculate the ACF. It will almost certainly be equal to zero or statistically insignificantly different. Well yes, there are no significant linear relationships between these series, all trivial, no fish.
But we Alexians don't count ACFs, we count non-linear dependencies - any dependencies. That's what the joint distribution of these two quantities is for. We have it. (By the way, it is also needed for calculating the ACF, which is something people usually overlook.)
The namesake immediately took the bull by the horns and counted the mutual information of these two series.
I evaluated the chi-squared test of the relationship between the two random variables.
The results are very similar.
- The joint distribution of the two quantities represents a surface. What, are we moving to another rayelnost?
We've been there for some time, it's just that not everyone realises it.
It's all just a fishing rod project so far, it's not a fish at all.
The only way to check the efficiency of the market is to check it yourself or else it will stay in the plane of "believe it or not. I can't say it any better than I already have:
I hope I'm wrong, but your "magic blots" are more in my head, the TS by channels is very similar to the TS by intuition, here's an unfulfilled forecast
I "twist" the levels with different builds, I see them working 50/50 in the proposed TS with very similar results, I suspect that even if I take a decade of price movement, they will coincide on the history or in the near future
Igor, I don't tend to idealise anyone or anything. But I already said
- These are not TSs, these are models of market movements. You have to build a TS based on them yourself.
- The author has ALWAYS been against forecasts. The essence of his work is expressed in one phrase - bounce, go to the previous one, breakthrough - go to the next one.
- As far as I can see it's not a prediction, but a possible target, which I personally have not focused on.
- You are looking for flaws without trying to get to the bottom of what's being proposed. I understand your scepticism, but you started off on the wrong foot. Try to get into the rules and the essence of the constructions first. To make it easier for you - Vadim's channel is the same candle, but without relation to the TF.
The TAdv postulates the movement development according to six control points. The channel is the points 1 and 2 in the Tadv. Swing is points 1,2,3 by Tadv. Note that neither Jan, who is one of the authors of TAdv, nor Vadim, the author of V-Channels and V-Swings, prove and convince anyone of anything, appear here only in strictly defined cases and do not ask anyone for anything, do not advertise anything. They do not flout and behave correctly. Isn't it an indicator of conviction and inner strength? They just unselfishly help and share their developments. TAdv was presented more than 10 years ago, VKanals and Vsvings, I'm afraid to lie, something like 7, have been tested by time and have a lot of followers. The only way to test the effectiveness is to delve into it yourself and test it to see if it works. Or else it will remain on the plane of "believe it or not".
Good luck.
Alexey! I, too, use mutual information to select meaningful inputs for a neural network. Usually I get models with predictability of returns of 52 - 54%. It seems to me that good results are obtained when mutual information is > 0.1 bit. For volatility it is possible to find such significant inputs.
Oh, good to see someone who has also learned how to apply TI to the problem of selecting significant variables come in.
Only your advice is a bit incomplete or something. The simple fact is that the average information of a data source can be different. The threshold meaningful mutual information will also depend on that. What is your average information flow H(X)?
On the five - hrenfx.
The posts are fine, quite on point. Besides just about applicability of TI to the subject, some have doubts here.
All right, let it be so, with fractality I overdid. There is, but it is not ideal. More precisely, fractal invariance is not ideal.
What statement - I do not understand.
Here it is clearer.
1. Two because we take a conditional source and a conditional receiver and try to find out if there is any dependence between them.
We make up an alphabet (I divide the distribution of returns into quantiles, it's more convenient for me; my namesake does it a little differently, but it doesn't affect the result much), apply TI. There are some doubts about the communication channel. Perhaps time is the communication channel.
If we fix the difference between the bars (the numbers are calculated in MT4) to be, say, 238, - then the source is the series
return(Bars-1), return(Bars-2), ... return(238) (about 80 000 values in 12 years)
The receiver is a series
return(Bars-1 - 238), return(Bars-2 - 238), ... return(0).
In short, just a series of returns and the same, shifted relative to itself by 238.
It is possible to calculate the ACF. It will almost certainly be equal to zero or differ from it statistically insignificantly. Well yes, there are no significant linear relationships between these series, all trivial, no fish.
But we Alexians don't count ACFs, we count non-linear dependencies - any dependencies. That's what the joint distribution of these two quantities is for. We have it. (By the way, it is also needed for calculating the ACF, just not everyone understands it!)
Tezka immediately took the bull by the horns and calculates the mutual information of these two series.
I evaluated a chi-squared test of the relationship between two random variables.
The results are very similar.
Yes we have been there for a long time, just not everyone understands it.
All this and there is so far only a draft rod, it is not a fish at all.
Thank you Alexey, it's all in the clear now.
In the context of this problem, the characteristics of the communication channel are absolutely irrelevant, they will be accounted for automatically via information entropy.
Hmmm, the experience was a success, my post was on the thread for about 5 minutes, but it was even blocked, and not by local regulars
What was the experience?
But we, Alexei, are not counting ACFs, but non-linear dependencies - any dependencies. That's what the joint distribution of the two quantities is for. We have it. (By the way, it is also needed for calculating the ACF.
Well said! We, Alexei, are in favour of market inefficiency. And we already have practical results showing this, but invisible through the prism of the classical statistical-econometric approach.