Econometrics: one step ahead forecast - page 114

New comment
 
faa1947:
No you***ing. Specifically. Compare matlab and shiryaev.

compare matlab and Shiryaev? Compare by the weight of the distribution kit?

So far I've seen only arbitration in Shiryaev.

I'm telling you that you don't read much, you mostly write and mostly the same thing :o)

 
faa1947:

I am not a teacher. I have been involved in investments all my life and, in parallel, I have sometimes given lectures.


Dear faa1947, how old are you, if that's not a secret? I'd also like to address you by your name and not by the serial number 1947. The inscription "SunSunich" obviously means Alexander Alexandrovich?
 
C-4:

Dear faa1947, how old are you if not a secret? And I would also like to address you by your name and not by the serial number 1947. The inscription "SunSunich" obviously means Alexander Alexandrovich?
Yes.
 
Farnsworth:

compare matlab and Shiryaev? Compare by the weight of the distribution kit?

I told you that you don't read much, you mostly write the same things :o)

Let's get back to our rams. We have one and very simple and universal.

We have a primitive regression model. It is shown that inside the sample it has a profit factor much larger than 10. Outside the sample it is a bit more than 1 and even that is doubtful. This model is "correctly" constructed.

Question: why does this "right" model not have the property of stability or predictability?

 
faa1947:
Yes.

You were born in 1947?
 
faa1947:

Back to our rams. One and very simple, but universal.

We have a primitive regression model. It is shown that inside the sample it has a profit factor much greater than 10. Outside the sample it is just over 1, and even that is questionable. This model is "correctly" constructed.

Question: why does this "right" model not have the property of stability or predictability?

(1) you haven't shown anything, that's the point.

(2) the fact that you have identified the model (not you, but Envil) does not mean anything, see item 3.

(3) the series is not stationary, the distribution and ACF are non-stationary (if you remember stationarity in the narrow and broad sense). The model parameters you will get will not be stable by definition, they will drift strongly. Moreover, for such series there is no notion of statistical sampling, matrix average, the size of samples does not determine anything.

(4) The parameters of the process your model "generates" do not correspond to the parameters of the original process. Simply, you will generate a totally different process, which has nothing to do with reality.

(5) further ... next see point 6

(6) "I remember you went to your topic several times. "That's it, I'm leaving. Don't worry, I'm barely interested in you anymore. But it was my fault, it was my natural curiosity, I looked and made sure that nothing has changed here so far, "I'm an econometrician and everyone else is **** :o)".

 
Vizard:

Sanych were you born in 1947?
Here we are looking at the nature of the problems, not the year of birth.
 
Farnsworth:(3) the series is not stationary, the distribution and ACF are non-stationary (if you remember stationarity in the narrow and broad senses). the model parameters that you will get will not be stable by definition, they will drift strongly. Moreover, for such series there is no notion of statistical sampling, mat average, sample size does not determine anything

If on point, it's good to see as in point 3

The original quotidian is unsteady and that's a fact.

We bite chunks out of it. The most obvious trend. Smooth, smooth and absolutely stationary as it is deterministic

We have the remainder - nonstationarity could not go anywhere and it is there - it is shown.

ACF shows that the trend was not eliminated completely at the first step. Again we mark the trend.

The residuals again. Again it is non-stationary. We check for ARCH and if so, we model it. I.e. simulate the kind of non-stationarity that we know how to do. Still more than nothing.

We will look at the residue. It is almost non-stationary. We are lucky. But most importantly, it is less than a pip. Let's spit on it. The error is too small.

Repeat your conclusion, but tied to a specific algorithm. It is implemented above with all calculations and graphs.

 
faa1947:
Here we are looking at the essence of the problem, not the year of birth.


ok...then let's not torture...the HP trend stands out...use a simple example to look at it step by step (bar by bar)... The trend that you take should not redraw! (the first bars on the right) otherwise all measurements are meaningless and wrong ...

Farnsworth's sampling in p3 was right... there's no way around it... but it doesn't have to be the same as in the software's help... and if you play with it, you can improve the cut... Although of course it's all nonsense by and large... nothing good can be predicted...

 

Повторите свой вывод но в привязке к конкретному алгоритму

Look, you're stubborn to the point of self-loathing.... one more time:

Choose as your series model, the incremental model: B(n)=B(n-1)+epsilon(n) (everything developed, developed for it) rather than B(n)=trend1()+trend2()+...trendp()+e. You have no idea about the trend patterns that sit in it and can never correctly identify them, especially as they change from time to time. Price is a multifractal, a very complex process

for your model to be applicable, you need to get (otherwise the model is easier to throw away)

  • a stationary distribution
  • a stationary ACF (or close to it)
  • statistical similarity of the model and source series (what you will predict). Once again, you are generating a series that has nothing to do with reality .

This is the minimum necessary (but not sufficient) With price this kind of thing doesn't work, try going to a transformation.

1...107108109110111112113114115116117118119120121...139
New comment