Author's dialogue. Alexander Smirnov. - page 3
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
2 khorosh: The indicator is indeed quite decent, but still I haven't found such a parameter method(smoothing method), in which it is by all parameters better than Djuric: it is almost always too frequent (fluctuating) on fluxes, though sometimes faster on jumps. Something close to Djuric is obtained at method=1.
It seems that Djuric has really made a very good adaptive filter.
LeoV, thank you. There are differences, but they really are very small. And the picture posted in the explanatory text probably corresponds to a very large value of the smoothing period (strong lag).
2 khorosh: The indicator is indeed quite decent, but still I haven't found such a parameter method (smoothing method), in which it is by all parameters better than Djuric: it is almost always too frequent (fluctuating) on fluxes, though sometimes faster on jumps. Something close to Djuric is obtained at method=1.
Looks like Dzurik really made a very good adaptive filter.
I'm kinda surprised to hear you say that, mathematician (because most people would believe your assertion). Tell me what is more adaptive? How do you measure the quality = figure and how do you calculate it ?
Give me the formula for what a TF should be adaptive to, then I can make a really adaptive TF and give it to you. (Juric would be worse, I think).
LeoV, thank you. There are differences, but they really are very small. And the picture posted in the explanatory text probably corresponds to a very large value of the smoothing period (strong lag).
LeoV, isn't it too difficult to visually compare legal and http://codebase. mql4.com/en/1356, which is the same name? Or is the legal one for Omega?
Here - two pictures. Period 14, phase 0. Very similar, by the way. Good algorithm for MT4. I can post it with some other period, if you want.
Decided to add my own picture
There is an Adaptive JMA. And at the intersection of regular JMA and Adaptive JMA you can already work on it. I just discovered this with surprise. .... All have a period of 14. Adaptive varies from 14 to 48.
I may have been too hasty about adaptivity, as I don't have a good idea what it actually is. I think it is just the ability to change the calculation algorithm depending on current conditions (sideways or trend activity). Different adaptive muwings apply different "flat/trend" criteria - fractal dimension, volatility, etc.
Djuric has four requirements which his filter must meet:
1. Minimum lag between signal and price, otherwise triggers come late.
2. Minimum overshoot, otherwise MA produces false price levels.
3 Minimum undershoot, otherwise time is lost waiting for convergence.
4. Maximum smoothness, except when price gaps to a new level.
Translation:
1. Minimum lag between signal and price; otherwise signal arrives too late.
2. Minimum overlap [something like Gibbs phenomenon - Mathemat]; otherwise the MA produces false price levels.
3. A minimum "underlap"; otherwise time is lost until the signal converges with the price.
4. Maximum smoothness - except for price gaps.
Juric & Co. solved the problem brilliantly and showed it on many examples (you don't even need to understand English, Prival: everything is explained at the level of a good comic book with vivid pictures). Of course, this does not mean that his filter can be bluntly applied as a replacement for muwings. The feedback from delighted users stresses several times that the signals should only be used in a certain context. But even with the bluntest use ("two muwings") there are still fewer false signals.
Kudos to you, Prival, for solving the daunting practical problem with Kalman for processing radar information, also quite noisy. But we are now really trying to understand exactly why JMA is so good on market data.
I want to reiterate that no most perfect and contrived adaptive filter like a muvinge with perfect characteristics will, by itself, solve the problem of creating a robust strategy. The problem is deeper than that; you know that from our private correspondence.
It's all very simple. The indicator has 2 inputs. One - for Close, the other - for indicator that shows trend of ADX type (or any other) and two parameters - minimum period and maximum. Minimum period - at minimum ADX, maximum - at maximum ADX. That's pretty much it.
LeoV
The question goes a little deeper. In order to answer that 1 indicator is more adaptive than another. You need to know what it should adapt to.
If we simply talk about the price. The most accurate (not lagging, not shaking, etc.) is Close[0]. But this is not good. We should remove what prevents us from finding the right direction (noise). And in order to answer this question correctly and correctly (from a mathematical point of view). It is necessary to answer the question of what is noise and what is signal. Only then we can say that some indicator adapts better to the useful component (signal) that moves the market.
And it is not very difficult for a good DSP specialist to make an optimal+adaptive indicator for a signal (models) that Djuric cites as a proof.
LeoV
The question goes a little deeper. In order to answer that 1 indicator is more adaptive than another. You need to know what it should adapt to.
It adapts, of course, to the trend. The "bigger and stronger" the trend - the longer the JMA period. And this, as I understand it, is correct... .
Here's a slightly idiotic model for you, Prival: if you consider the returns (signal increments), the signal is zero, the noise is a random process with a p.d.f. of the Cauchy distribution type and an ACF, which you know empirically. There are no measurement and quantization errors. Of course, the price as a result of integration will jump around the expected payoff, because the tails are very thick and dependent.
The model is extremely rigid, perhaps even tougher than the market itself. But if your filter will work on such a model, it will work anywhere.
P.S. By the way, Djuric made such a suggestion: if one of those who bought his creation provides a filter that works better than his (by four criteria described above) on Cauchy-type data, he'll sort of get his money back. And this is just an unambiguous hint at the noise model, which he himself was guided by.