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
A model without overfitting should produce stationary residuals regardless of sampling
Why should it? There is a whole class - adaptive models, where to go?
My model is not a landmark, its lifetime is one bar, so to speak a one-bar product. I don't believe in models that can live for years. The reasons you have described and they are widely known.
... then we can talk about the stationarity of the residuals produced by the model.
I'm not interested in the theory of stationarity. I'm interested in the model. There are at least three basic questions:
1. Are there any extra variables in the model?
2. Should additional variables be included?
3. Is the model construction process completed?
Stability is a criterion for stopping the construction of the model. That's all. Next, forecast by one bar. Where is the looking ahead here?
And to build a model that will feed until retirement is pure communism, a great and bright utopia.
So this criterion is simply not sufficient. There is something you are not considering.
How do you have confidence in the sufficiency of the model, if you just typed in a few dozen necessary tests (and even their necessity is not obvious!), hoping that someday this necessity will turn out to be sufficient?
Why should you believe in them? They are there.
So this criterion is simply not sufficient. Something you are not taking into account.
How do you have confidence in the sufficiency of the model if you have just typed in a few dozen necessary tests (and even their necessity is not obvious!), hoping that someday this necessity will turn out to be sufficient?
Sufficiency is like the end of geography. If there is no autoregression in the residual and modeled ARCH (if there was a need), then there is nothing to model. The knowledge is over.
You can, but they don't. Give me an example of an indicator whose text is accompanied by R-squared. Indicators are used and it is not known to what extent they reflect the quotire or if they reflect at all. Judge by eye, "of course a great indicator"
do...just don't write... judging by eye - haven't seen any great indicator... I don't need to analyse its maths to do so...
It really is. We have almost a stable residual. Shift the window by 1 bar and we have to change the model parameters (number of lags). This can be clearly seen in the table by the two outermost columns, where the number of lags is shown.
It's called one word - fit to the story...
Give me a link to a proof of the claim that these conditions are sufficient for prediction.
I don't remember the proof, but it applies everywhere. I will give my (someone else's) rationale. Once again: kotir = trend + noise + periodicity + outliers. From this I take trend + noise. Reversibility is present: by adding trend + noise we get kotir.
What do we know? The answer is obvious - trend. Apart from that, it is meaningless to analyse noise while there is a trend in it - it will score the statistic characteristics of noise. We should model trends until there are no trends left in the noise. When all the trends have been identified (I have not seen more than two levels), then there is ARCH in the noise. If there is, then we also know how to model - modelled. Is the residual stationary? Fine. We do not know how to model further. Not being able as a sign of sufficiency.
I remembered though. The stationary residual may have the property that the probability of changing the sign of the increment is higher than the probability of keeping the sign.
PS. Sad if the stationary residual is large in scope. Ideal when less than a pip.
You can't remember them because they don't exist. It would be too easy to make money in the market then...
do... just don't write... judging by eye - haven't seen a single great indicator... I don't need to analyse its maths to do that...
It really does. Got a near stable residual. Shift the window by 1 bar and we have to change the model parameters (number of lags). You can see it clearly in the table by the two outermost columns which show lag count.
it's called one word - fitting to the story...