Hearst index - page 11
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
Is there a built-in Hearst function in Excel? If so, please name it. >> thank you.
Hearst calculation is done in a script. In Excel you just need to logarithm and find the straight line.
The Hearst calculation is done in a script. In Excel you just need to logarithm and find the straight line.
absolutely right.
the function is called TIP(y,x)
The Hearst calculation is done in a script. All you need to do in Excel is to logarithm and find a straight line.
Then you can do it in MQL, here's 'Useful functions from KimIV' to help you.
I will just test Hirst for myself. I have been thinking about refining Spirmen for a long time but I still have not got around to it. Maybe the synthesis of these two indicators is just what I need.
Trend (global) and Hurst have nothing to do with each other, Sergei. Hurst shows, roughly speaking, the ability to microtrend. I.e. Hurst index says something about microstructure of time series, but not about a trend. It seems that with H >> 0.5 (closer to 1) something can be done on a time series (profit) - just because it is not martingale (differences of neighboring samples are correlated). And non-martingale - because adjacent samples are dependent.
I'll show you pictures, though you must have seen them yourself. They all are from the Peters' "Fractal Analysis...". Note that there is no trend anywhere. Hearst's numbers are 0.72 (upper left), 0.76 (upper right), about 0.9 (lower left) and well under 0.5 (lower right). You know what a Wiener process looks like (H=0.5).
It's all a qualitative picture too, of course.
That is, the Hurst index says something about the microstructure of the time series, but not about the trend.
It's all a qualitative picture too, of course.I will add my thoughts on this, if I may.
A fairly complete characterisation of BP is given by the Autoregressive model. In general terms, BP can be thought of as the sum of a deterministic component and a random (noise) component:
This is the AR model for price increments dX. With its help, knowing the p-value of previous increments, we can predict the expected movement of the quotient with a known certainty. Moving from the price increments to the forecast of the price itself is not difficult; to do so, just add the expected price increment to the last value of the instrument price and you will get a price forecast for the next step.
Above I showed identity of Hearst Ratio (HR) calculated for each TF of the quote and correlation coefficient (CC) between neighboring readings in a series of the first difference of the quote (stochastic BP is shown in red, EURGBP min is shown in blue). The coincidence can be considered satisfactory and even in favor of the CC - smoother dependence, other things being equal, and incomparably simpler expression for the calculations compared to the PC:
There is, however, a fundamental difference. The PC is a deeper and more complete characteristic of BP in comparison to QC, because it assesses the quotient as it is - in its entirety, with all its internal links and features, without resorting to artificial separation of features. QC in these conditions exploits the only parameter available for its analysis - the relationship between neighboring counts of cotier increments, and that's it. The fact that the results coincide indicates only the weak correlation of the long-range counts (in fact, the second left count has almost no effect on the future value of the instrument's price increment) with the expected movement. In fact, the opposite can happen (deep links appear) and QC will fail, while the PC will work correctly.
This is the similarity and the main difference between these two methods of BP analysis.
It should be emphasized that PC is an integral characteristic of BP, which says nothing about the specific properties of the relationship between the incremental counts. In contrast, the AR model is quantile in full and gives a quantitative characteristic of these relationships (coefficients in front of dX under the sum sign), which allows us to exploit them 100%. But there are also limitations due to the linearity of the approach used. AR-models that take into account non-linear relations between increments have more complete information. But again, this model should be developed by us and it is not the fact that it is optimal.
And this is where Neural Networks come in... Non-linearities form their basis, and the ability to learn gives them the necessary flexibility.
And that's where Neural Networks come in... Non-linearities are at their core, and the ability to learn gives them the flexibility they need.
No one is arguing that, but for persistent and antipersistent BP or sections of BP, the trade
The tactics are diametrically opposed, so the NS has to learn to take the PC into account when trading.
Maybe it's better to feed her ready-made than to wait for her to learn to see it herself.
Maybe it's better to feed her ready-made than to wait for her to learn to see it for herself.
It's a matter of debate as to what's best. What are the criteria for judging whether it's better?
You are appealing to the PC as the ultimate truth, but it is just a tool, which has its own possibilities and limitations. And it is not a fact that to wait until HC itself reveals a feature is worse or more expensive than to feed it something visible, but not the best. Besides, in the process of searching, NS is focused on maximization of profit (speed of account growth), and PC is focused on persistence of BP, which still should be somehow tied to TC and only then to account growth.
Trend (global) and Hurst have nothing to do with each other, Sergei. Hurst shows, roughly speaking, the ability to microtrends. I.e. Hurst index says something about microstructure of time series, but not about a trend. It seems that with H >> 0.5 (closer to 1) something can be done on a time series (profit) - just because it is not martingale (differences of neighboring samples are correlated). And non-martingale - because adjacent samples are dependent.
I'll show you pictures, though you must have seen them yourself. They all are from the Peters' "Fractal Analysis...". Note that there is no trend anywhere. Hearst's values are 0.72 (upper left), 0.76 (upper right), about 0.9 (lower left) and well under 0.5 (lower right). You know what a Wiener process looks like (H=0.5).
It's all a quality picture too, of course.
Some free time has appeared. I will try to program it and post it here. I will use Matcad to make all of them with explanations of where and how I modeled them.
My aim is not to get some good quality pictures, but to investigate the Hearst's exponent, its performance with different input signals (on test models) and on this basis to understand its performance field and possibility to use it.
Here are the models. If you think you need some more, write it down.
I have some free time now. I will try to program everything and post it here. I will do everything in Matcad, with explanations of where and how I modelled it.
Purpose: to get not good pictures, but to study Hearst's exponent and its work with different input signals (on test models),
How are you going to get the Hearst figure for the current situation? It means to consider a limited number of N bars at the moment in order to calculate Hearst on this particular sample. So you need another criterion to find the moment in the past, from which the calculations for the current moment are made.
And that's where Rosh hit the mark. To calculate the Hearst figure you need a lot of historical data. It is not a muwing whose memory is limited to a period, but a global characteristic of BP as a whole - or a large chunk of it.