Regression equation - page 7
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
P.S. Your thread threatens to become one of the most fascinating and informative on this forum. There are few of these, really few.
Indeed, so many luminaries in one place. So, what are we discussing? Regression, I think.
There is a button on the terminal called " Linear Regression Channel". You don't like it, although it's very nice and handy for obliques.
I don't think I like it either, although I sometimes draw it - for the sake of cuteness. The reason is simple: by drawing a linear regression we proceed from the assumption that the quotient model will be a set of segments. Complete nonsense.
What if we take a different order of regression? That means that the quotient model will be a regression of that chosen order. Also nonsense.
So maybe first verbally define what happens in a quotient, then pick a model, estimate that model and that's when the regression issue is resolved.
The chewed ARPSS(p,d.q) model - contains in particular regression with order p. If one reads Box, one finds that p is not necessarily greater than 0, i.e. there are BPs where the BP regression does not apply at all and is modelled by other means.
Without verbal assumptions about BP, i.e. assumptions about the kind of model we want to identify, the conversation is purely cognitive, mathematical and TC irrelevant.
Tell me, to which data do you intend to apply regression analysis? And to what end?
Nothing personal, but you get the impression that you want to eat soup with a hammer because it nails better than a spoon.
I just do not get the idea, why do you want to apply regression analysis? And for what purpose?
Candid's example is a good one.
Throw-ins - take it as data and marvel at the fat fosts...
Hilarious! (c) Kolobok with a fat cigar.
There is a culture in life. The same purity of thought and skill must be in the analysis.
(A careful statistician would have thrown out this observation - and would not have reveled in the fourth order of his "pulled by the ears crookedness" and the result of "proof" which was foreshadowed) - it is a pity that Сandid did not mention the author.
Love the hammer, though. :)
hrenfx- Way to go! After all, we all remember that in the beginning there was an idea... the model, and then the approximation.
So here too. if we know that there are measurement errors (they are inevitable in the physical-real world, AND THEN it's easier for us to accept them as the basis of deviations) - one model.
But if we take a subscription to the data (e.g. the favourite DC quote stream from the venerable Morning Star... :) - then the errors will be quite different, and the market model all the more so.
But not in the observations, but in the assessment of the situation of the players' errors. And their fumbling - in no bar data will you notice. Prival is right - but the fumbling itself is fumbling.
And elipsoids with NP-complexity of solving problems of extremum search with constraints - an ornament to the table. Like there are such problems. To the topic of the topic of the topic - like Golokhvastov's speech. He is bragging himself!
Imho.
-- The child has spilled out.
To a normal, in the sense, applicable to forex ;), discussion of regressions - I call!
And don't forget about liquidity - the engine of progress!
;)
I just don't get the idea, so why do you want to apply regression analysis? And for what purpose?
What's there to grasp. Ask others why they want to apply Machka, RSI, MACD and other ancient classic indicators. But here it is regression.
I understand faa's arguments, but only partly. The flow of quotes does look like poorly glued pieces of plots, which are not badly predicted.
P.S. Michael Andreevich, well let's not muddy the waters about the software of the company, which owns the resource, eh?
FreeLance: (a careful statistician would throw away this observation - and not revel in the fourth order of his "pulled by the ears curvatures" and the result of the "evidence" that was predicted)
All this statistical tambourine dancing around rejecting "random" large spikes is in fact a veiled attempt to force the real distribution (real life distribution with fat tails) to a fine-tailed one. All the same desire to work with convenient formulas rather than inconvenient reality.
...And their fumbling - you won't notice it in any bar data. Prival is right - but the shara-ha-ha-ha-ha himself...
Here I am, just sometimes I can't write for reasons beyond my control :-))
And when they talk about models and fat tails. I keep thinking of Kamal and Kniff and reread their posts (sorry they are not on the forum very literate) good thread. mathematician even "called" me a Bengal tiger there :-))
https://www.mql5.com/ru/forum/105771/page15
kamal 09.12.2007 00:50
Well in the end, in order not to play here only as an "ideas killer" I will express a very simple idea, which I used to push even in my article here at mql4.ru, and which as I grew in practical trading experience, has become more and more important: the standard Gaussian model of geometrical random walks is saved from all problems by rethinking only one parameter: time. This idea has already been mentioned here, but it's not a sin to repeat it again: look at the tickframe! And the effects like "heavy tails", like "volatility", and many other things will disappear.
alsu 21.09.2010 21:44
everything is already formalized, read the link, the one in Russian (first on 3 pp). The quantile regression problem is reduced to linear programming problem: find minimum of linear function under linear constraints.
Write formulas. It takes 2-3 minutes to program and find a solution in matcad. A lot of people here can do it just fine...
alsu 21.09.2010 21:44
write the formulas. It takes 2-3 minutes to program and find a solution in Matcadet. A lot of people here can do it just fine...
What's wrong with forum pictures? I can't insert png.
Read http://www.nsu.ru/ef/tsy/ecmr/quantile/quantile.pd f, paragraph 2, the very beginning, where the LP problem is described. All the formulas are there.
What is there to catch. We have already learned that the traders are not willing to apply Machka, RSI, MACD and other stuff from the classic indicators. Well, here it is regression.
The "want to" approach seems wrong. I'll give you my view:
1. A transformation is performed over the market data (the simplest one is to take the tops of the ZigZag. For example, taking all ticks as they are - these are tops of the ZigZag with the condition of a min knee of 1 point). We obtained a data matrix, where each column corresponds to an observed parameter (e.g., the price of some financial instrument) of the market. And each row is a vector of market condition in the observation space.
2. An assumption is made that if we find an effective regression (linear, polynomial or any other - it does not matter), it will give relatively low deviations at some interval of data outside (before and after) its construction.
3. Statistically investigate the behaviour of the regression outside its plots. Weak spots are found. Find the causes.
As most people do (all the same points) :
1. one (two or three at most) financial instrument is taken. No transformation is carried out, they just take all ticks (bars - more often).
2. Anything is used, for "want" reasons.
3. There is no statistical research. An Expert Advisor is written and optimized hoping hopefully.
I understand faa's arguments, but only partly. The flow of quotes does look like poorly glued pieces of plots that are not badly predicted.
My arguments, as I think and would have liked, were deeper.
The TS always has some kind of model underneath it - regardless of the desire of the author of the TS. If the author of the TC denies this fact - then the chukcha riding on the tundra with the song: what I see, I sing. This is not an insult - Chukchi for thousands of years have been driving from point A to B without any landmarks and problems. So do the TC authors - they get to the point of profit.
If you think about the model, you can learn a lot of unexpected results about your trading system, not visible at first glance.
The most accurate model is Prival's with its ticks. But is it possible to make a one-hour forecast based on ticks? On an hourly chart it is in principle possible - there is only one candle. And on ticks? Forecasting at least 60 candles forward? It is possible? On a non-stationary market? And what is the confidence interval? My understanding is that there is no answer to these questions.
Let's take the ARPSS model (p, d, q) where d=q=0. If p=1, it is a straight line. This section of the market is quite possible. Moreover, a prediction is possible. But we would have to accept the conditions external to this model that: there is a trend and it is stationary, and the noise should be less than SL. If there are no other elements in our model that bring it closer to real life, we will very soon find out that the deposit is null.
Which model is better? A more complex one that takes everything into account like Prival's or a rougher one? There is no theoretical answer. There is an empirical one - in case of identical model estimations, the simpler one should be used. From the latter follows a fundamental point about models: there is nothing to argue about models if you don't have a sound system for evaluating them.
Given the above, I conclude about the topic: no model with regression built in and no evaluation of the result obtained. Schoolboy talk on regression, and in a very primitive way (remember the post on regression analysis).
FreeLance 22.09.2010 04:28
But if you take a subscription to the data (e.g. the favourite DC quote stream from the ever-memorable Morning Star... :) -Then the errors will be quite different and the market model even more so.
The model is decided before the DC and the DC cannot influence the model. Initially, you look at the quote and try to see: is there a trend or not? Are there cycles or not? And what is the volatility? Different DCs for the same instrument and timeframe cannot affect the answers to these questions. The pattern is ahead, everything else is behind.
it's already formalized, read the link, the one in Russian (the first one on 3 pages). Quantile regression problem is reduced to linear programming problem: find minimum of linear function under linear constraints.
I was thinking here, gradient descent will work worse than simplex-method, since grad-t is more general. All other things being equal, it is not less iterative.