Econometrics: one step ahead forecast - page 50

 
joo:

From my engineering practice.

A colleague of mine was once sent on a business trip. He spent two years developing a vibro-loader. A vibro-loader is a device having something like an eccentric, and it was designed for sinking piles into the ground.

So, he left on a business trip with his rattling wonder. Our clients call from there: "Your specialist came, installed his rig on the pile and said - this (bleep) won't work! He took out a bottle of vodka, drained it in two gulps, and disappeared in an unknown direction." .....

The man didn't admit to the end that his work was shit. But one day he did.

I asked you to be specific on the topic of the thread!
 
faa1947:

You can make up all sorts of things.

I originally laid out my verbal description of kotir = trend + noise. This description makes sense in terms of prediction, as the trend is predicted.

In this thread I have raised a very narrow issue: a 1 step forward forecast. I have proposed a model and am trying to find out if the forecast can be trusted. If you can, why, and if not, why not. On this topic I would like to hear opinions and suggestions. And willing to do the dirty work of coding to test hypotheses. This is what I call specificity.


The main question is which market characteristic tends to persist. For example, linear regression - the trend is linear and persists for some time. There are other characteristics and accordingly models. Your HP also makes an assumption about the conservation properties of certain characteristics. But any model cannot objectively reflect the market all the time - one needs filtering. When one or another model is adequate to the real market.
 
faa1947:

Here is part of the summary table:

What to change?


I do not use such characteristics. The traditional type is the dependence of the target on changes in model/TC parameters. And much depends on understanding the model - what process it uses effectively and what makes sense to use and analyse for it and what contradicts it. I.e. don't dig with an escalator all over the place, but with a trowel where needed))
 
Avals:

I.e. not digging everything with an escalator, but with a trowel where it's needed.)
Well, it's not all over the place. It speaks to the proper design of the model. For example, ARCH. The table shows that the probability of no ARCH is always greater than 10%. But this is already the result of changing the number of lags in the regression. I.e. a selection is made and we are sure that there is no heteroscedasticity in the residual. This model is more "correct" than the one with heteroscedasticity. Among such correct models we should look for profitable ones, if we can do it.
 
Avals:

The main question is which market characteristic tends to persist. For example, linear regression - trending is linear and persists for some time. There are other characteristics and accordingly models. Your HP also makes an assumption about the conservation properties of certain characteristics. But any model cannot objectively reflect the market all the time - one needs filtering. When one or another model is adequate to the real market.

I understand this and am hoping to get such a set of non-similar, "orthogonal" models.

At the moment I have linear and non-linear models available to me in regressions. Thinking the bottleneck is the highlighting of the trend.

 
faa1947:

I understand this and am hoping to get such a set of non-similar, 'orthogonal' models.

At the moment I have linear and non-linear models available to me in regressions. Thinking the bottleneck is isolating the trend.

 


Explanation.

The results are calculated in pips and observations, which means: one trade - one observation. A total of forty bars. Every day one trade - either long or reverse and vice versa.

Profit within the sample. We take a sample of 40 bars. For these 40 bars, the regression is estimated and then the algorithm starts to make the forecast beginning from 1 bar of this sample.

Profit outside the sample. It takes 40 bars, estimates the regression and then makes a forecast for the next bar out of sample.

I think that profit in observations is more accurate, because it does not depend on the value of increments

 

Here's the thing... Even at this stage:

this "persistence" of the error should be alarming - it's some kind of hint that the process is "inanimate".

.

This echoes one of the points of identification theory, which demands sufficient spectral diversity of the signal under study. It's a duality.

 
avtomat:

Here's the thing... Even at this stage:

this "persistence" of the error should be alarming - it's some kind of hint that the process is "inanimate".

.

This echoes one of the major points of identification theory, which demands a sufficient spectral diversity of the signal under study. It's a duality.

This is the purpose of model building in econometrics.
 
faa1947:
This is the purpose of model building in econometrics.
I'm not quite sure what exactly is the objective?