The market is a controlled dynamic system. - page 18

 
Vizard:

thanks interesting...but the question is why Hedrick and not MA for example? or SSA...

You could say that HP is of higher quality than MA, but that's not the point. The whole point of decomposing an initial non-stationary quotient is to get a stationary residual.

and how much does the residual depend on the number of samples? i.e. 137 now and if we take 2000 for example.

I've tried it, the prediction error increases. I think we need to reduce the sample size. We need a one step ahead forecast, don't we ? I don't see the problem with overfitting as for the forecast for the next candlestick we estimate the model again. The coefficient is likely to change. I would like to keep the model structurally unchanged.

I suggest we finish, respecting the author of the thread. I plan to open an appropriate branch. I invite you there.

 
faa1947: ... the whole point of decomposing the original non-stationary quotient is to get a stationary residual...
It is the opposite, in econometrics the point of decomposition is to maximise the stationary trends, thereby reducing the error introduced by the non-stationary residual. I.e. the more non-stationary the residual, the better the decomposition is produced. In general terms this is the case.
 
-Aleksey-:
On the contrary, in econometrics the point of decomposition is to isolate the stationary trends as much as possible, thereby reducing the error introduced by the non-stationary residual. That is, the more non-stationary the residual, the better the decomposition is produced. In general terms this is the case.

thereby reducing the error introduced by the non-stationary residual.

This is news to me. As long as the residual is non-stationary - no prediction is possible, because in a non-stationary series the variable is mo and variance. This is the point of using ARCH models, which model different kinds of non-stationarity in the residual.

 
Vizard:


I don't get it. The average error is 1190 pips, or is it in what units. R squared is ridiculous, but how do you understand the negative value?
 
Vizard:

If yes - the fall is not caught either... But it is interesting... We should use TS and look at equity...

That would make more sense.

.

There are signals in the figure that relate to the current moment - so don't pay attention to them.

The moment in question is highlighted with vertical lines - this is what we are interested in.

Thus,

on TF D -- OpenSell ........... on TF H4 -- CloseSell

Here options of downwards work only -- Sell

The rebound is obviously false - the variant Buy is not even considered.

 
-Aleksey-:
In econometrics, the point of decomposition is to isolate stationary trends as much as possible
NR is about isolating trends, but the aim is different - it is about getting a stationary residual
 
faa1947:


I suggest we finish, respecting the author of the thread. I plan to open an appropriate thread. I invite you there.

Well, why don't... Go ahead, please. I too am very interested to see the possibilities of econometric methods.
 
faa1947:

thereby reducing the error introduced by the unsteady residual.

This is news to me. As long as the residual is unsteady, no prediction is possible, as the variable in the unsteady series is mo and variance. This is the point of using ARCH models, which model different kinds of non-stationarity in the residual.

The prediction is made on the extracted stationary components. That's why they are highlighted. The residual is considered an error, although it can also be predicted if desired.
 
Vizard:

I forgot to clarify ... What parameter is Hedrick used with ? i.e. how many bars are subsequently redrawn in the pricing model ? or are they cut off immediately ?
I wrote, lambda = 10. There is no problem with re-rating as the forecast is one step ahead, the next step will take a fact, the model is recalculated and the forecast again. The model is defined by a formula on the last two bars. HP is not propagandizing, just taken because of fame.
 
-Aleksey-:
The prediction is made on the basis of the allocated stationary components. That is why they are highlighted. The residual is considered an error, although it can also be predicted if desired.
The error must be stationary, otherwise it is not defined on the forecast one step ahead and may be arbitrary, although everything is fine on the sample.