Econometrics: one step ahead forecast - page 124

New comment
 
Farnsworth 10.01.2012 11:34

then prove the existence of a trend in the quotes, beforehand, remembering to define it clearly.

Trend to me is a quality thing. We look at the handsome ZZ and see a trend. Yesterday, today and tomorrow and the day after tomorrow.

I don't have a trend problem, I have a smoothing subtraction residual problem, which has an analytical expression that is more than stationary and which I am going to extrapolate beyond the sample. There are plenty of smart guys like me on the forum, but unlike them I say: the problem is in the residual, because there is non-stationarity which can spoil any beauty. And I deal with the remainder. This is the step-by-step decomposition of the original quotient.




 
You are free to profess your beauty. But your beauty is lifeless. Static. A sculpture. A stone. Time easily kills such beauty. I think beauty is in evolution. The ability to change... and change... This is the only way LIFE can pave its way into the future.
 

faa1947: Вся наука именно такая: решает то, что видит, а из того что видит, то что может, а из того что может, то что можно применить.

I'm sure you see the first point ("decides what he sees") incompletely.

 
Mathemat:

I'm sure you see the first point ("decides what he sees") incompletely.

Everyone sees everything incompletely. the absolute truth is not available to us.
 
DDFedor:
You are free to profess your beauty. But your beauty is lifeless. Static. Sculpture. A stone. Time easily kills such beauty. I think beauty is in evolution. The ability to change... and change... This is the only way LIFE can make its way to the future.
I do not profess to be static; on the contrary. I say, let's identify the different characteristics of the quotient, the most unpleasant among them being non-stationarity. Let's identify it and work with it and not bury our heads in the sand.
 
faa1947:
Farnsworth 10.01.2012 11:34

then prove the existence of a trend in the quotes, beforehand, remembering to define it clearly.

Trend to me is a quality thing. We look at the handsome ZZ and see a trend. Yesterday, today and tomorrow and the day after.


(1)

Here's where I suggested an example of random rambling(https://www.mql5.com/ru/forum/136555/page80), if you want, you can target a ZZ and the handsome trends will appear in huge numbers. Are you going to predict them the same way?

(2)

ZZ itself is not predictable, neither is a quote. And it's even worse with GZ, if various tests confirm the existence of non-linear relationships in the initial process, they disappear completely in GZ. No surprise there, GZ are the places where the price was actually lowest. Try from the "contrary", build a GZ based on the maximum "concentration" of that price.

(3)

I don't have a trend problem, I have a smoothing subtraction residual problem, which has an analytical expression that is more than stationary and which I'm going to extrapolate beyond the sample. There are plenty of smart guys like me on the forum, but unlike them I say: the problem is in the residual, because there is non-stationarity which can spoil any beauty. And I deal with the remainder. This is the step-by-step decomposition of the original quotient.

Then, options like this:

  • you need more advanced models that can "correct themselves", like ARIMA/ARMA/... They have a second component that "shifts" the result in the right direction based on error analysis. Better yet, take neuronics, it is much more promising in this sense. Actually, AR is a neuron :o)
  • Stabilize the basic "simple" model and study behaviour of its parameters in a non-stationary environment. Only after that will you understand what to do with the remainder

 
Farnsworth:


Here I have suggested an example of random rambling(https://www.mql5.com/ru/forum/136555/page80), you can target ZZ if you want and handsome trends will appear in huge numbers. Will you predict them the same way?

I saw an article stating that it is impossible to distinguish a stochastic trend from a deterministic one - we all cannot do it, that is why we ignore it.

ZZ in itself is not predictive, nor is a quote.

They do and will continue to exist. The fact that we don't know what's on the right side of the screen is our problem, not the indicator's.

We need more advanced models that are able to "correct themselves", like ARIMA/ARMA/

The model used in this thread is ARIMA with second degree of integration and variable number of lags in AR and MA

Stabilise a basic "simple" model and study the behaviour of its parameters in a non-stationary environment. Only after that will you understand what to do with the residual

Studied, but it's not clear what to study. My models have a bunch of properties (test results). There was an idea that if all the tests were passed ("correct model") there would be a prediction. But that didn't happen. The whole topic stopped at this point.

 
faa1947: I say let's identify the different characteristics of the quoter, the most unpleasant among them being non-stationarity. Let's identify it and work with it and not bury our heads in the sand.

You are fixated on quotient alone. You may differentiate it (take the differences) as much as you wish, even 10 times. But what good is it, where is any guarantee that it will continue to be the same in the future? Where are these guarantees in the model itself (along with their estimates)?

Stability (resilience), including to the future, can be sought in other price functions, not just in the quotation itself. To do that, you will have to use your brain and stop chiselling quotes in only one way (by chart regressions).

And please format quotations so that you can see who is quoting and who is replying. I'm already confused. At least put arrows (>>) in place of quotes, if you can not get the standard.

faa: There was an idea that if all the tests are passed ("correct model") there will be a prediction.

Where did this idea come from? Why were you so sure of this?

 

ЗЗ - это доказательство наличия трендов - они есть и будут

Why, I like your persistence already.

I saw an article that said it's impossible to distinguish a stochastic trend from a deterministic one

Exactly, hence the need for other approaches.

The model used in the thread is ARIMA with the second degree of integration and variable number of lags in AR and MA

that's not enough, do you want to describe the market with the second degree of integration? If you want at least some approximation of the same AR model to the market, the order should be at least 200-300

Studied, but it's not clear what to study. My models have a bunch of properties (test results). There was an idea that if all the tests passed ("correct model") then there would be a prediction. But that didn't happen. The whole topic stopped at this point.

How to what?

(1) Take a fixed sample length for which you are sure EViews identifies the model correctly.

(2) Traverse the sliding window in the direction of "astronomical" time with one bar (counting) step

(3) For each such step (let there be at least 100, taking into account manual work) let EViews determine the optimal coefficients of the selected model

(4) Write it all down in a table: step number, coefficient values, error/residual, criteria

And look at it all together, how it all changes.

 
faa1947, price increments depend on a huge number of factors, some of which are not even related to previous prices. Whichever method you apply, even in an ideal case it only takes into account a small fraction of them. Most of the time, their influence on the quote is insignificant - at the level of noise - and quite rarely they come to the forefront and are able to create some movement. Then there is an opportunity to match them using certain cause-effect relationships. So, don't initially build a model that is always trying to predict something - filter the bazaar. And build subject-oriented models, not just what's in your software. What processes may be behind pricing based on the basics of trading. Speculative, investment, conversion, etc. Certain assumptions about the mass behaviour of traders, building a model on this basis, verification (maybe even based on ecometrics).
1...117118119120121122123124125126127128129130131...139
New comment