Econometrics: why co-integration is needed - page 15

 

"believe - disbelieve" -- this is the field of religion, the subject of research of theology (from Greek theologia, from theos - god and logos - word, doctrine) -- i.e. the science of god: )))))

;)))) what a crooked co-integrated econometric path you have taken!!!

 
Mathemat:

I see your point.

But I don't believe in static arbitrage.


It is shown above that the delta is too small. So it is not a matter of faith, but a specific calculation. Generally there is a defect in the idea: we are looking for mismatch between those going in the same direction. We should be looking for mismatch when they go in different directions.

I'm leaning towards the opinion of the anonymous person who wrote about portfolios. There, as part of the more expansive idea of an efficient portfolio, cointegration is probably useful.

For speculators the usefulness of cointegration is not clear to me. Well, except to justify testing. Something new, but so far very weak evidence of usefulness.

 
If bank yields " 10 per cent, any sustainable strategy with sustainable returns above bank yields should attract the attention of investors.
 
faa1947:

But where is the constructiveness?

Good question, but for the sake of objectivity you should ask yourself that question from time to time.

Still, I'm sure that your model outlined on the first page is not very workable and:

... says that you can use this regression for prediction without worrying - it has a stationary residual.

If I were you, I'd be a little worried. From a number of indications, you've just "crushed" it, and there's not much predictive value in it yet.

You haven't answered the question - what is your stationarity?

 
Farnsworth:,


and there is not much predictive value in it yet.

There is no prognostic value in it.

You didn't answer the question - what is your stationarity?

A series is stationary if the mean and autocovariance do not depend on time.

So far we collectively found out that cointegration is used in arbitrage, but I calculated that the delta is too small, so it is questionable to use in arbitrage.

More interesting to me is the assumption that we can trust the test if the quotient is cointegrated with the balance from TC and cannot be trusted if not cointegrated. I asked collective to give initial data for analysis for hypothesis testing - only tara gave such information. I have posted the result. There is a lot of information in the championship, but was not able to copy, no one helped to do it.

 
faa1947:

and there is not much predictive value in it yet.

The objective of forecasting has not been set

If you are forecasting, then clearly the objective as such, in the interpretation of forecasting, has been set. Another thing is that you do not yet understand what to do with the resulting curve, but that is a whole other part of the system and setting.

A series is stationary if the mean and autocovariance are independent of time.

Not exactly so, if you take the theory, albeit not strictly. Stationarity is interpreted in a broad and narrow sense. Not "average" but if the distribution and ACF are stationary. In the narrow sense you must prove that you have such a distribution and the parameters of this distribution (not only the mean, it may not exist) remain the same during the whole process. More than sure that your ACF is not stationary at all, which means that your whole atrolabe is guaranteed not to work, i.e. you can't even properly predict, let alone use it.

So far collectively we have found out that cointegration is used in arbitrage, but I calculate that the delta is too low, so it is questionable to use in arbitrage.

Oh, it's everyone's choice of belief :o)

More interesting for me the assumption that one can trust testing if the quotient is cointegrated with the balance from TC, and cannot be trusted if not cointegrated. Asked collective to give initial data for analysis to test hypothesis - only tara gave such information. I have posted the result. There is a lot of information in the championship, but could not copy, no one helped to do it.

On this forum there is a colleague HideYourRichess I quarrel with him from time to time, but there are places of "agreement", respectively, different ways to this understanding came (I used fractal analysis). So the test of any balance curve for correctness is very simple, the less "diffusion" and more "linearity" (in quotes I wrote, it needs deciphering) of this curve - the more reliable result and it can be trusted. And I got that fractal parameters of this curve should lie in certain ranges. One can make an analogy, also just from fractal analysis. Sounds or music is often classified simply, into "like" and "dislike" and it is not so important blues or jazz. So, 'like' turns out to have its own 'fractal boundaries', and a sound getting into this range begins to 'like'. I may have explained it vaguely, but that's not the point.

And in essence, the task of your TC is to convert a perfectly curved quotient into preferably straight line with positive coefficient (and it seems to be what avtomat is trying to convey in his "sketches"). And this is where I'm not really clear about what cointegration of TC and kotir you're talking about. OK, let's say you want to cointegrate the balance curve and kotir by getting a third curve. What would that get you?

 
Farnsworth:

If you are forecasting, then clearly the objective as such, in the interpretation of forecasting, is set. Another thing is that you do not yet understand what to do with the resulting curve, but that is a whole other part of the system and setting.

Not really, if you still take the theory albeit not strictly. Stationarity is interpreted in a broad and narrow sense. Not "average" but if the distribution and ACF are stationary. In the narrow sense you must prove that you have such a distribution and the parameters of this distribution (not only the mean, it may not exist) remain the same during the whole process. More than sure that your ACF is not stationary at all, which means your whole atrolabe is guaranteed not to work, i.e. you can't even properly predict, let alone use it.

Oh, it's everyone's choice of faith :o)

I have a colleague HideYourRichess on this forum, though I quarrel with him from time to time, but there are places of "agreement", respectively, different ways to this understanding (I used fractal analysis). So the test of any balance curve for correctness is very simple, the less "diffusion" and more "linearity" (in quotes I wrote, requires deciphering) of this curve - the more reliable result and it can be trusted. And I got that fractal parameters of this curve should lie in certain ranges. One can make an analogy, also just from fractal analysis. Sounds or music is often classified simply, into "like" and "dislike" and it is not so important blues or jazz. So, 'like' turns out to have its own 'fractal boundaries', and a sound getting into this range begins to 'like'. I may have explained it vaguely, but that's not the point.

And in essence, the task of your TC is to convert a perfectly curved quotient into preferably straight line with positive coefficient (and it seems to be what avtomat is trying to convey in his "sketches"). And this is where I'm not really clear about what cointegration of TC and kotir you're talking about. OK, let's say you want to cointegrate the balance curve and kotir by getting a third curve. What does that get you?

I sit inside EViews and trust the tool rather than my own understanding of the terms. This enables me to use a ready-made product instead of reading an insane number of books, often with questionable content. Moreover, in the end, everything fits together and I always have enough tools which have been proven to work many times over.

The cointegration I'm testing :

I test the unit root of the original quotes

I select a vector that, when subtracting one quotient from another, gives a stationary quotient in the residue (unit root test).


Okay, suppose you want to cointegrate the balance curve and the quotient by getting a third curve. What does that give you?

And this is the hypothesis. If the two series are cointegrated, i.e. the difference between them is stationary, then the test can be trusted and it does not matter whether it is positive, negative, straight balance line or curve.

If not cointegrated, the test cannot be trusted. It has to be tested. I wanted to do it experimentally. For the tara data this was confirmed. The result is above.

 

to faa

Я сижу внутри EViews м доверяю этому инструменту, а не собственному пониманию терминов. Это дает мне возможность использовать готовый продукт вместо чтения безумного кол-ва книг, зачастую сомнительного содержания. Причем в конечном итоге у меня все стыкуется и всегда хватает инструментов много кратно проверенной работоспособности.

yes it is just a tool and the falsity of interpretation of estimated statistics and data will not be ensured "automatically". It depends entirely on the analyst.

And this is the hypothesis. If the two series are cointegrated, i.e. the difference between them is stationary, the test can be trusted and it does not matter if it is positive, negative, straight balance line or curve. If not cointegrated, the test cannot be trusted. It has to be tested. I wanted to do it experimentally. For tara's data it proved to be true. The result is above.

I think it is more of an illusion. How are you going to co-integrate? Let me guess - assign it to EW, and what it will do is pull it down by the ears within almost any model and you will get a fake stationarity. There are no criteria here, any profit curve can be cointegrated with a quote (any) by picking a model. What does it give you? Where will you understand that when you "optimise" (which is what you will be doing) you will need to stop and how will you separate the parameters into bad/good?

PS: still try to complicate the model, make it more sensitive or something. Your model "squeezes" the process, in fact - changes the scale very much. As a result, the process "steps" in such small increments that it just can't get very far.

REQUEST: Specially highlighted.

Generate your new process with a length of say 3000 counts. Take the first 1000 counts and the last 1000 counts. There will be another 1000 in between. And post it here for the first and outermost ACF segments. And we'll all look at your "stationarity" together with the naked eye

 
Farnsworth:

to faa

...

HOWEVER: Specially highlighted.

Generate your new process with a length of say 3000 counts. Take the first 1000 counts and the last 1000 counts. There will be another 1000 in between. And post it here for the first and outermost ACF segments. And we'll all look at your "stationarity" together with the naked eye

Yes, I almost forgot - I'm interested in the increments, but for the company you can also show ACF of the source (on the same graph for each species, it would be more convenient)

A little addition: ACF is possible for the first 100-300 samples, probably more will not be needed.

 
Farnsworth:

Yes, I almost forgot - increments, I am interested in them, but for the company you can also ACF source (on one graph for each species, so it would be more convenient)

A little addition: ACF is possible for the first 100-300 samples, probably more will not be needed.

Stationarity is checked with a unit root test. Existing subtleties that occur due to ACF are solved within the test or the choice of test type (there are several at my disposal). I see no reason not to use existing achievements and start repeating what was done 20 years ago e.g. by Hamilton