Statistics as a way of looking into the future! - 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
I don't understand why you are against time in the model. you can't take time out of the model. The forecast from Friday to Monday is different from Thursday to Friday, it's different horizon (different time of forecast). And there are factors to take into account when market opens and closes, it's not because of time which we are used to, but it (time) is directly related to it. This rule is incorrect to make forecasts without saying at what time interval it will be done.
For example: +10 pips in a minute, and +10 pips in a year :-)
You see, let's assume, in the red, that you have wrongly assumed that this theory is suitable for predicting price. And I don't think it is at all. I want to uncover that. So I ask why all of a sudden this theory? If the premise is there (apart from similarities in the name and purpose of the task at hand), then I would be wrong. So I'm already in the discussion.
Right, I definitely sympathise with you! Here I am too, all the time I ask such questions to Prival: well, why has he decided that the laws of the market fit into the system of Newtonian differential equations? Why, all of a sudden, the price should be like an aeroplane on the radar screen and move like a massive body under the action of a forcing force? Let him give the rationale for his approach... So far, he hasn't. Just, pretends not to notice (doesn't understand) and only cites pretty pictures and questions about the Matrix.
As for me, I know that in 1950s Kolmogorov proved the theorem that any functional dependence of n variables can be approximated with given accuracy by a set of elementary weighted adders. All that remains is to find the weights of the inputs of these adders. You see? In this formulation, I don't need to know the type of the function dependence and I don't need to draw a hyperplane in my mind! It is enough to ground the algorithm of finding adders' weights and they will do it and serve me on a silver platter. This is a fundamental point, this is the main difference from autoregressive models and the like. I don't need a model, if there is one, it will be found, even if it's not stationary (just retrain the NS).
This is what determined my choice in favour of NS and the direction of my research.
No, well let's predict the weather outside. Only how will you then generate trading signals based on these forecasts?
+5
I, for example, don't see much point in coaching complex mechanisms, including NS, on indicators that are essentially price transformations. What's the point in that? Give to the NS the same data as to be inputted to the indicator and the NS will adapt to the indicator functionality. So why bother it with extra data?
bstone, imagine an ordinary poker... if you look at it from a certain angle, you can see a straight stick and never guess what it really is. The task of pre-preparation of data fed to the NS inputs, is to simplify as much as possible the work it will have to do on the input data, and present it in the most digestible form for it. And this is very important! It needs, as Gogol's Viy, to lift its eyelids and show the object. Ha, you will say! - Everything has been done for her, so what is the point of it? And you'd be wrong. She needs to be fed a restaurant menu. It's an aristocrat of mathematics.
Oh, what persistence. I gather that your, Vita, knowledge of dynamical systems theory is extremely modest. Otherwise you would know that dynamical systems theory allows you to express even such complex and inherently chaotic systems as self-organisation.
Well let's go back to the basics first. What is a system as understood by the aforementioned theory? A system is any object of nature whose state changes in time according to a certain law. If the market is not such a system, then as it has been rightly pointed out - we have nothing to do here. But we are good optimists, aren't we?
By a dynamic system is meant a system, which state is uniquely determined by initial conditions and time. In such a form it makes no sense to pull it on the market and no one here, I hope, does it.
However, certain classes of dynamic systems are perfectly suited to modelling chaotic systems capable of self-organisation. And if one can skillfully solve the problem of identifying the parameters of suitable dynamical systems, they can successfully play the role of a model of the chaotic system under study.
Now imagine that there are techniques for passing from the phase space of the original system to the phase space of auxiliary systems that can be analyzed by existing methods.
My humble knowledge allows me to see my bolded blah-blah-blah about the magical properties of the theory, but does not allow me to see that price has anything to do with that theory.
Mendel's laws can predict too, but they don't apply to price as I understand it. How does the nonsense you wrote become applicable to price? I'm not discounting either Mendel's laws or Dynamical Systems Theory, but why did you choose to use Dynamical Systems Theory rather than Mendel's laws?
Personally, I believe that comprehensive and all-encompassing "if skilfully" begins with the question of the applicability of this or that theory to price. Does the spell "there are methodologies" automatically make a theory suitable for predicting price?
Apart from pointing out that I am unprofessional, is there any other argument that "certain classes of dynamical systems are perfectly suitable for modelling price", because price has exactly the properties (here you list the properties) that theory allows for? Can you specify what assumptions need to be made in the input of the theory for these "separate classes of dynamical systems?"
No, well, let's predict the weather outside. Only how will you then generate trading signals based on these forecasts? - You're twisting my words. I see your point. I do not want to get away from my vested interest - what price and some prediction theory have in common. So let's leave that point for later.
For example, I don't see much sense in training complex mechanisms, including NS, on indicators that are essentially transformations of price. What's the point in that? Give to the NS the same data as to be inputted to the indicator and the NS will adapt to the indicator functionality. So why bother it with unnecessary data? - Exactly, I didn't mean indicators either.
About all those rebounds with probability distributions I haven't seen anything useful that could work on the market yet, out of models based on pure statistics. Why do you think there are so many such models: AR, ARM, ARMA, GARCH, EGARCH... the list goes on to several dozen. They simply do not work, although they solve a much simpler task - volatility prediction. - Oh, yes, they do! They do, but they only predict the volatility of the portfolio return. They do, because they assume a normal distribution of returns for that portfolio, which is true only theoretically, with new caveats about the independence of instruments. But to predict price with these models, one has to agree that price has something in common with the normal distribution law, or, more strictly, with parametric statistics. That's why I doubt that a method suitable for modelling portfolio volatility is suitable for predicting price - because of the inapplicable assumption into which price has to be shoved.
Really, I definitely like you! I, too, keep asking Prival these questions: why would he think that the laws of the market fit into a system of Newtonian differential equations? Why, all of a sudden, the price should be like an aeroplane on the radar screen and move like a massive body under the action of a forcing force? Let him give the rationale for his approach... So far, he hasn't. Just, pretends not to notice (doesn't understand) and only gives pretty pictures and questions about the Matrix.- That's right, and here everyone keeps silent, like a matter of course, that the price is ready to jump into any theory, as long as it is embraced there by skilful hands.
As for me, I know that in 50th of the last century, Kolmogorov proved theorem that any functional dependence of n variables can be approximated with given accuracy by set of elementary weighted adders. - My gut feeling is that we are talking about any functional dependence with certain parameters, even if we don't know them. Price has no such property, it does not fit in the bed of parametric statistics, so I doubt that any functional dependence exists. Unfortunately parametric statistics is only strong where the system has certain parameters, only then does it give us beautiful results.
bstone, imagine an ordinary poker... if you look at it from a certain angle, you can see a straight stick and never guess what it really is. The task of pre-preparation of data fed to the NS inputs, is to simplify as much as possible the work it will have to do on the input data, and present it in the most digestible form for it. And this is very important! Like Gogol's Viy, it needs to lift its eyelids and show the object. Ha, you will say! - Everything has been done for her, so what is the point of it? And you'd be wrong. She needs to be fed a restaurant menu. It's an aristocrat of mathematics.
This is what I cannot agree with. The well-studied property of NS that makes their application so attractive in quite different fields is their ability to learn and subsequently successfully approximate nonlinear dependence of any complexity.
What I can agree with is that if we train a network by inputting let's say the previous open price difference and two indicators that use the price of previous 50 bars in their calculations, the NS will show better results than the ones obtained by the NS the input of which only the previous open price difference. But in fact, if we train such a network by inputting the previous 50 quotes, then in theory it should learn the combined dependence of output from inputs, which takes place when applying the indicators.
However, it is obvious that it is technically much more difficult to train a network with 50 inputs than one with 3 inputs. But this does not mean that indicators are useful by themselves. They are just crutches that help to avoid technical difficulties, but in the end they significantly reduce the capabilities of the NS. Isn't it so?
My humble knowledge allows me to see my bold blah-blah-blah about the magical properties of the theory, but does not allow me to see that price has anything to do with this theory.
For fuck's sake! How else can you ask that? I've already said that the market is a system. Imagine that the prices of all instruments traded on the market are the parameters of this system. And they all evolve according to some unknown law. Now do you see what the price has to do with the systems theory?
Dynamical systems theory, not Mendel's laws?
Apart from pointing out my professional ineptitude, are there any other arguments in favour of the fact that "certain classes of dynamical systems are perfectly suitable for modelling price", because price has just such properties (here you list the properties), which the theory admits? Can you indicate what assumptions need to be made in the input of the theory for these "separate classes of dynamical systems?"
Well, I'll say it again for the third time. Systems theory applies to the market because the market is a system whose parameters (prices) evolve according to some law. That doesn't mean it gives you the answer to all questions, but if there's a coherent theory that fits the system in question, why not use it? Or is it better to reinvent bicycles, point your finger in the sky and predict the weather?
Vita писал(а) >>
I don't want to get away from my vested interest in what price and some sort of prediction theory have in common. So let's leave that point for later.
That's great! They do, but only by predicting the volatility of portfolio returns. They do, because they assume a normal distribution of those portfolio returns, which is only theoretically true, with new caveats about the independence of the instruments. But in order to predict price with these models, one has to accept that price has something in common with the normal distribution law, or, more strictly, with parametric statistics. That is why I doubt that a method suitable for modelling portfolio volatility is suitable for predicting price - because of the inapplicable assumption into which price has to be shoved.
There you go. You didn't get into the essence of this question in sufficient depth either. Exactly all of these models "sort of" work because they take into account the fact that volatility does not fit within a standard normal distribution. About portfolio returns - that's not relevant at all. The volatility forecast has nothing to do with portfolio returns and their distribution. Another thing is that volatility forecasts are mainly used to assess the risk of a portfolio, but that's another story.
No, no, we're talking about any one!
but about functional dependence - i.e. a parametric law. However, that is not even the point. Why should we assume that there is any functional dependence behind the price ? There is no assumption. Just a belief in the mechanistic nature of the world and envy of Nostrodamus' laurels.