Obtaining a stationary BP from a price BP - page 14
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
Well the DFT generates 2 coefficient arrays for sines and cosines + the average value of Ak0. Since we are using DFT on each sample, Ak0 is actually an LMA with period = window size. Correspondingly, we need to extrapolate the muwings in order to reconstruct harmonics around them
I used cosine transform there, you can use Hartley transform (their coefficients also have stationarity, Hartley also has formula of transition to Fourier transform and back). It seems to me it is necessary to predict them separately, which increases the error. But maybe I don't understand something.
What are you!!!!! Look at your avatar from the outside - God forbid you should dream about it. I just want to set you straight. That's not a correct definition. A more correct one is given, for example, on this site in the TA section. There are simply established definitions which do not need to be changed and tucked away with who knows what. Moreover none of the tools (including yours, presented on this mite) do not fit the definition in this fucking wikipendia (there is simply no real analysis of price behaviour and moreover no regularities, which it describes). Anything that has to do with price falls under this definition at all. For example, FA (yes, yes, it does), perfectly self-sufficient stochastic financial mathematics, described for example by Shiryaev in two volumes (facts and models), "stochastic control systems with fixed/random structure" which are also self-sufficient. All of the above work with price series, but work on completely different principles than TA.
If you define TA by methods of analysis, I agree. I simply defined TA by the object of analysis. That's all. And to say that one is right and the other is wrong is strange. They are both correct, but they describe the subject from different points of view. You know, like in a relational database - one object - many relationships.
No, no, no! not the pinched ego! By the way, I never said anything about your knowledge or called you names ... :о) And how do I know what your subconscious is hiding?
... you look pretty pathetic. Knowledge and understanding of the subject is negligible - I have all the moves written down!!! ))) But forget it ... why bother. I'm just that - for historical truth.
If it didn't matter - I would have forgotten about it long time ago.
It REALLY doesn't matter. The essence of what we do does not depend on the subtleties of classification. It's just one way, one way, another way. It's all just words.
GOES!!!! AGREED!!! NOT ANOTHER WORD!!! I'm generally peaceful, maybe a bit nerdy, but peaceful. :о))))
Peace.
Hooray. And that's it - the subject is closed.
If you define TA by methods of analysis, I agree. I simply defined TA by the object of analysis.
And as always, regardless - common sense wins out. :о)
The problem with adaptive TS is that they are also retrained according to some algorithm, which is built into them, but it may not coincide with the algorithm of market changes. That is, the algorithm of market changes can coincide with the algorithm of TS retraining at some period of time, but then it can "go away". The market doesn't change according to a given algorithm - that's the problem.....
I agree with you.
This is the reason why it seems sensible to move towards single-parameter adaptation algorithms. Then, the number of handles that can be handled is the minimum possible. It may seem that in this marginal case, the efficiency of the model is not the highest and a more "advanced" result can be obtained by tuning 2,3 or 5 parameters. However in the conditions of weak stationarity (on the edge of its absence, which are price BPs) the optimal configuration will be the minimum, because its requirements for the length of the training sample and stationarity of the optimized parameter are minimal. It is impossible to prove strictly this statement, but its completeness concerning the market is proved by experience and justice for the limiting case when stationarity tends to zero.
From this point of view, it does not matter which algorithm to use for BP price analysis (MACs, oscillators, discrete breakdown, etc.), it is important that the number of optimization parameters is 1 (2 - in the extreme case). Obviously for single-parameter algorithm we need to decide on parameter choice (there are only 2 of them - price and time). In my opinion, the more important parameter is the price scale. It is by looking at price changes that we make decisions about entering or exiting the market, and only secondly we analyze time (I don't mean pipsewise strategies). For a two-parameter model both parameters must be considered - price and time, but here, as I mentioned above, we can encounter the problem of "short" stationarity (the characteristic time of existence of which is shorter than the minimum length of the training sample). Three or more parametric models, it seems to me, makes no sense to consider at all, as they are a linear combination of these two parameters.
Separate interest is presented by analysis of the EMA (it is the smoothest muving of all with the minimum FP). This MA contains one parameter, and takes into account the amplitude and time components of BP. Two in one, in short. Although, it's possible that with EMA it's all a hoax and no miracle will happen.
grasn wrote >>
Hi Sergei. :о) Glad to see you. Where have you been, what interesting things you studied?
>> Wait, let's take it one step at a time. We have a problem of series transformation, with given quite specific properties:
(1) stationarity
(2) normality.
(3) possibility of reverse recovery
All this time I was investigating the problem of searching for the optimal length of the training sample and its relations with characteristic time of quasi-stationarity of processes on the market. It turns out that the required value is at the level of 5-10%. Then it has to be retrained. The brokerage company commission in its turn defines the minimum size of the price movement, and a gradual but sure increase of market efficiency with the growth of the trade horizon defines the operation area unambiguously. And I kind of settled on that.
As for answering your questions regarding BP conversions, I'm not at all aware of the feasibility of such a conversion. Your ". it will allow to use standard methods of stat-processing..." doesn't say anything. Please state the concept itself.
This is accurate, but obtain a transformation of the original price series that has the following properties
is quite possible, of course with some acceptable assumptions. To the question "why do we need it" the answer is very simple and the only one - it is an opportunity to use the tried and tested framework, and no more. My imho.
It's impossible to argue with that.
And as always, regardless - common sense wins out. :о)
I agree with you.
It is for this reason that it seems sensible to move towards single-parameter adaptation algorithms. Then, the number of handles that can be handled is the minimum possible. It may seem that in this marginal case, the efficiency of the model is not the highest and a tuning of 2,3 or 5 parameters may produce more "advanced" results. However in the conditions of weak stationarity (on the edge of its absence, and price BPs are exactly of that kind) the optimal configuration will be the minimum one, as its requirements for the length of the training sample and stationarity of the optimized parameter are minimal. It is impossible to prove strictly this statement, but its completeness concerning the market is proved by experience and justice for the limiting case when stationarity tends to zero.
From this point of view, it does not matter which algorithm to use for BP price analysis (MACs, oscillators, discrete decomposition, etc.), it is important that the number of optimization parameters is 1 (2 - as a last resort). Obviously for single-parameter algorithm we need to decide on parameter choice (there are only 2 of them - price and time). In my opinion, the more important parameter is the price scale. It is by looking at price changes that we make decisions about entering or exiting the market, and only secondly we analyze time (I don't mean pipsewise strategies). For two-parameter model both parameters must be considered - price and time, but here, as I mentioned above, we can encounter the problem of "short" stationarity (the characteristic time of existence of which is shorter than the minimum length of the training sample). Three or more parametric models, it seems to me, makes no sense to consider at all, as they are a linear combination of these two parameters.
Separate interest is presented by analysis of the EMA (it is the smoothest muving of all with the minimum FP). This MA contains one parameter, and takes into account the amplitude and time components of BP. Two in one, in short. Although, it is not excluded that with EMA it is a cheat and a miracle will not work.
I investigate the problem of searching for the optimal length of the training sample and its relation to the characteristic time of quasi-stationarity of processes on the market. It turns out that the required value is at the level of 5-10%. Then it needs to be retrained. The brokerage company commission in its turn defines the minimum size of the price movement, and a gradual but sure increase of market efficiency with the growth of the trade horizon defines the operation field unambiguously. And I kind of figured it out.
As for the answer to your questions about BP transformation, I'm not aware of the reasonability of such transformation. Your ". it will allow us to use standard stats processing methods ..." does not mean anything. Please sound out the concept itself.
1) Weak stationarity - this is not an edge of its absence, and the markets are not. weak stationary process - is when the mean = const, and st.dev. depends on time (but not strongly) due to the heterogeneity of the markets and the variability of volatility, which is inertial - this is a predictable process.This dependence can be easily described by a formula, that allows to determine the price extremum in any definite period of time, beyond which the price will not go with the probability, determined by us; do not confuse with the determination of price movement direction.Such a weak-stationary process can be thrown out to the garbage heap. The task is to obtain a highly profitable and controllable process that sacrifices extra stability.
2) In order not to run into the problem of "short" stationarity there are FA and capital pricing structures and statistical tests.
3) EMA is a common IIR filter with FZ=1/3 my opinion, even if it is 1/10 - it won't change anything, and the bars themselves already have FZ.
4)The feasibility of such a conversion is high, but the lamppost with the "forex" sign needs to be changed.
4)The feasibility of such conversion is high, but we have to change the lamppost with a "forex" sign.
This is perhaps a valuable observation, but the feasibility of such a conversion is still not fully revealed.
FOXXXi, think about it: we have a price series which is an integrated CB with nearly zero MO and non-stationary moments. The idea of converting it to a stationary series implies some functional dependence of moments on something else (e.g. time of day, etc.). The problem of "residualization", thus, is reduced to identification of this functional dependence and its exploitation... and what if this dependence does not exist or it is non-stationary itself!
We are trying to build a sandcastle, ignoring the fact that it will crumble sooner or later anyway. I suspect that we fall for beautiful words and science, losing sight of the impracticality of all these actions. This is the kind of game we have, in which the result is not important, but the process itself is interesting. And what will this "stationarity" give us? We don't have to optimise our Expert Advisor every time we need it. It's a great challenge to reoptimize it once a month! Anyway, based on my understanding of the problem, it is not worth a bite. Just like many other forex trading strategies.
The problem with adaptive TS is that they are also retrained according to some algorithm, which is built into them, but it may not coincide with the algorithm of market changes. That is, the algorithm of market changes can coincide with the algorithm of TS retraining at some period of time, but then it can "go away". The market doesn't change according to a given algorithm - that's the problem.....
Good point. But sometimes you can set the adaptive EA in the right direction by initial optimization and it will go further by itself. But then the market and the adaptor have different opinions.
FOXXXi писал(а) >>
The objective is to have a highly profitable controlled process and somewhere to compromise with super stability.
In this case, the process is highly risky and uncontrollable, i.e. totally unsteady.