FR H-Volatility - page 35

 
NorthernWind, Better himself (in the sense of his strategy) - hardly discussed. It's not too spread out, and there are only hypotheses about PNN, inputs and outputs. There was an attempt, 'Probabilistic neural networks, packages and algorithms for MT4' but it is more related to different software. See his profile thread, there are some bits of information there.
 
Mathemat:
Yurixx, I have already written my preliminary conclusions about equivolume bars here: 'Need an indicator reflecting price in operating time.'. It is not quite what you ask, and the discovery did not make my mood better either as I was expecting something closer to Gaussian.

But it does give me some hope - if, say, one throws it on this graph ... well, okay, I'll have to check it out...

And about the distribution of maxima, I think kamal gave you the idea, as far as I remember.


I saw that post when you wrote it. Noted for myself that you also got the results that Neutron and I got earlier. Noted also that you were going to sort out the High-Low stats for the equivolume bars, have been waiting for the results ever since. will they be ?

I had no illusions that switching to Equibar would make a revolution. If it were possible, this revolution would have taken place a long time ago - the idea lies on the surface. But the fact that it will make it possible to reveal any subtleties is quite conceivable. And it is often in the subtleties that the clues lie.

You're absolutely right, "that's not exactly what I asked". However, if you know an approach that will solve the problem, please share, pls. In my post on the last page, in the two quotes cited, the problem is formulated quite clearly, in my opinion.

By the way, pay attention to this thread 'Classification of trading systems and estimation of their value'. You don't seem to have been there yet, and without the problem you are trying to solve, all reasoning on that topic is constructing formulas from the ceiling.

 
Yurixx:

Noted also that you are going to sort out the High-Low stats for the equivolume bars, have been waiting for the results since then. Will they be ?

...

You're absolutely right, "that's not exactly what I was asking about". However, if you know an approach that will solve the problem at hand, then share, pls. In my post on the last page, in the two quotes cited, the problem is formulated quite clearly, in my opinion.

...

By the way, pay attention to this thread 'Classification of trading systems and estimation of their value'. You don't seem to have been there yet, and without the problem you're trying to solve, all reasoning on that topic is constructing formulas from the ceiling.

1. Of course they will. Just let me write it properly, this indicator. Hi-Lo statistics can easily be obtained by importing the necessary figures into Excel.

2. No, I don't know of such an approach. I haven't set out to get such a distribution theoretically yet.

3. I've been there and showed the author of this branch what absurdity his formulas show when applied to Better and TeamSky Expert Advisors. The main priority (and the most critical indicator) must be surely the strategy stability, not the profit factor (which Better has, by the way, is far from being ideal, but people are still fond of it).

If we talk about the super-task of strategy evaluation with synthetics, stability E could be measured, for example, depending on the estimate of probabilityp of killing the system on a certain time interval T for a given maximum drawdown D.

It is very similar that, all else being equal, p( k * T, D ) ~ k * p( T, D ) , i.e. the probability of killing the system on a 10-year interval is 10 times higher than that on an annual interval. Simpler is the same: the probability of survival of a system over k years is 1 - ( 1 - p ) ^ k ~ k * p for small p. So the test interval can be made standard (e.g. one year).

The dependence of this function on D is already much more individual. But here too we can try to approximate the dependence of the probability of death on D and again give some standard value of D - e.g. 20%.

Of course, this formula can be applied only if the preconditions are met - small enough p, decent expected payoff, etc. OK, let's say we have established what p equals for the system at standard T (year) and D (0.2). That would be P.

Well, the value of the system V may involve, say, our P and the recovery factor. And in general it is enough just to understand how a potential investor makes a decision. He assesses the risks and possible profits and makes a decision based on these figures.
 
Mathemat:

3. I've been there and showed the author the absurdity of his formulas when applied to Better and TeamSky Expert Advisors. The main priority (and the most critical indicator) should certainly be the strategy stability, not the profit factor (which, by the way, is far from ideal in Better, but people are delighted with it).


Ah, yes, I saw it, I read it. Sorry, forgot it was your post. :-)

And a potential investor makes a decision based on the risk-return ratio. That said, zero risk (i.e. the risk associated in the US with government liabilities) already has a corresponding return - bond yields. In principle there are also reference points for the yields of riskier instruments. But how the price of the instrument will depend on the increase in yields while maintaining the same measure of risk is the question. Imagine that you estimated the probability of losing a deposit by an MTS-company at p. At the same risk p , the average return of available market instruments is d. And MTS provides a return of 10*d. What should its price be ?

 

Hi all!

For the sake of interest I decided to generate a small NS in MatLab and see how it all works.

As input parameters, decided to use equidistant ZZ, let's try to predict its vertices. Equidistant chosen for reasons of maximal compression of input information with minimal losses of useful information.

The figures illustrate what I have said. The upper left shows the usual phase superimposed on the tick history, and the right one - its equidistant representation. I was not interested in prediction of the absolute price value, but only in prediction of the expected increment (we trade the price change), see the left figure below. On the right side is an example of NS truncation on an arbitrary sample of 100 values. You can see that NS has learned to predict the tops of ZZ absolutely accurately!

I must admit that I was a bit surprised to see results of step-by-step prediction without retraining in the area that the network "didn't see" during training (fig. below left), and an attempt of Wiener-type BP prediction (right):

There it is - either I'm wrong and inexperienced to have "looked into the future", or one of the two! For me this result was unexpected and impressive.

But, the fact that the prediction, on the unchanged parameters of the NS, of the Wiener series has failed, gives me a certain optimism.

 

Let's try to complicate the problem for NS.

Predicting the distance between the nodes in the zone is not a good thing, because this value, though unpredictable, has an average value equal to the double step of partitioning the zone. The NS can "detect" this and give a "trivial" prediction +-H at each step. This is not interesting (we already know about it). But we do not know where and to what extent the price will move after the next peak is formed! For this purpose let us build the series of increments ZZ minus2H . The red color shows increments in WP and the black color shows increments in transactions (TP) or price moves after the formation of the next peak in WP.

Pastukhov's paper gives an integral estimation (H-volatility) from which the expected direction and value for PT can be found. Unfortunately, this value is inside spread and is of no practical interest for trading. What is of interest is the forecast at each PT step with the accuracy that will allow moving out of the spread gates. Below (left) shows the forecast made by NS for a series of TPs based on EUR/JPY ticks, on the right - for a Wiener BP.

The result looks encouraging even now! It seems that this direction requires more attention.

 
The last two pictures, left and right?
 

Hi Sergei !

As far as I understood, you were feeding a sequence of segment size values to the network input. Or am I mistaken? If not, I wonder how many values you fed. Your NS is too good at predicting the next segment. Or was it just a coincidence? It's hard to call 5 points a statistically valid number.

 

Hi Yura!

Yes, I've inputted segment size first (prediction result shown) and then tried to predict segment minus 2H, this variant does not provoke NS to predict the sign-variance of WP sides (result shown). I gave 4 segments as input and predicted the next one, then, repeated the operation considering the new segment but without over-optimization and so on 5-10 times. Usually NS worsened its prediction ability after the 5th forecast and over-optimization was needed. The graphs are presented as my choice. I can't provide statistics, because I use a ready-made NS implemented in Matlab - press the button - see the result :-)

 

Well, that's impressive. As a first result - a quick sketch and a go - it's very good. It's not for nothing that I like ZZ. You can't give it to the input so directly...

And what NS (if it's no secret) did you build for it?