Econometrics: one step ahead forecast - page 100
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
You can discuss it in another thread. Are there any results (report, monitoring)?
That's for sure - there's not much to discuss in this thread...
Well, there is a result, of course. If there was no result, I wouldn't lie. Calculations are made based on the results, the model is corrected, technical parameters are determined.
How about this: I keep doing my experiment -- 10 months ahead -- if any questions arise along the way, I'll explain, I'll explain... But, agree, explaining what a derivative is would be too much... That is, some basic knowledge is assumed.
Yes, in order to understand it, you must put some effort, and sometimes considerable effort, and do some cognitive work. But I could not do that work of reasoning for someone, even if I wanted to.
Oleg, I don't need to explain what a derivative is. I am familiar with TAU to an extent sufficient for understanding what you have drawn. So about the basic baggage of knowledge - it is certainly not to me.
I simply pointed out that I didn't see any specifics (a model of something) in that thread. And I'm not the only one.
It's just bare schemes (see my previous post, I also asked questions there, which you did not answer).
Oleg, I don't need to explain what a derivative is. I know enough about TAU to understand what you have drawn. So about the basic knowledge, it is definitely not for me.
I simply pointed out that I didn't see any specifics (a model of something) in that thread. And I'm not the only one.
It's just bare schemes (see my previous post, I also asked questions there that you didn't answer).
No, no... about the derivative - it's not for you... it's a figurative statement... although the recipient didn't hesitate to show up :)))))
Well, naked is naked... To me, however, they have a very definite meaning.
Well... Let's leave it at that...
That's for sure - there's not much to discuss in this thread...
Well, there is a result, of course. If there was no result, I wouldn't lie. Based on the results, calculations are made, the model is corrected and technical parameters are determined.
And where are they (stats, monitoring)?
Generally, some individuals have an "interesting" logic -- if the result is shit, it's accepted as the norm....
Let's just say it's not suspicious. But the 70,000% is very suspicious. It's even odd.
That's because there are no gentlemen around.
Let's just say it's not suspicious. The 70,000%, on the other hand, is highly suspicious. It's even kind of weird.
ded and more promised ;)
I would like to return the topic to its roots.
Background information on the topic is presented in two articles:
Analysis of Statistical Characteristics of Indicators
Econometrics: One Step Forward Prediction
The following model has been proposed:
EURUSD hp1(-1 to -2) hp1_d(-1 to -1) eq1_hp2(-1 to -3) eq1_hp2_d(-1 to -4)
where hp1 is Hodrick-Prescott indicator from 1/DX, where DX is dollar indicator.
hp1_d - residue = 1/DX - hp1
eq1_hp2 is the Hodrick-Prescott indicator of the residual = 1/DX - (hp1(-1 to -2) + hp1_d(-1 to -1))
eq1_hp2_d is residual from previous smoothing.
lags (previous bars) are specified in parentheses. I.e. the model uses values of 2, 1, 3 and 4 bars respectively.
I made a one-week forecast using this model which showed positive results of 5 to 2.
Then I posted the test results in EViews. I am repeating that result here:
This table shows the properties of the model:
R-squared- the quality of the model's fit to the quotient, if = 1, it matches
S.E. of regression - the error of fitting the regression to the quotient. If we take 4 decimal places, the error ranges from 11 to 55 pips.
LM ACF - shows the probability of absence of autoregression in the residual. In red, where we cannot reject the hypothesis that there is no autocorrelation, i.e. there is one
Next two bars: tests for the presence of heteroscedasticity in the residual from the model. Shows the probability of absence. The table shows the results of optimization and modelling of heteroscedasticity where necessary, i.e. we can't see if it was initially present in the residual.
RESET test - probability of no specification errors: missing variables, functional form error, correlation with error (with residual)
Max Prob C is the maximum probability that the regression equation coefficients are equal to zero.
Lambda H1 to H2 is the lambda value for the Hodrick-Prescott indicator.
the last two bars are the number of lags in the model. Adaptation was applied and we can see that a shift of one bar leads to a change in the number of lags. Selection criteria were: min LM ACF and min Prob C
The summary results are as follows:
We have surprising results inside the sample and more than modest outside the sample. The value of the profit factor of 1.22 should not encourage us, as it can squeeze out a particular quotier movement. More objective is the profit factor in observations = 0.77, which shows that out of 40 trades (a trade on each bar) 22 were loss-making and 17 were profitable.
What is the problem?
A common approach to TS: create a TS, run it in the tester, get a bad result - modify it. What to change - do not know.
Idea:
Is it possible to find some properties of TS which could be a benchmark and by the values of which one could judge the quality of TS before testing it. I.e. - we test only a "good" TS. I'm sure the tester can show good results for an inoperable TS.
I suggest that you stop the ignoramuses' squeal that econometrics is bad and only because of this you should not pay attention to it. There are specific tools available. They are shown in the table. What can be squeezed out of specific tools for the construction of the "correct" TS? Let's remember: before you drive a car, you have to make sure that it is in good working order. We have a different attitude to TC - the serviceability is determined by the depot's drain.
Still ready to implement suggestions in EViews and post results.
faa, as I understand you have 40 forecasts and 40 results for these forecasts in the dry run. Could you make a separate table for clarity - number of n/a | cotir-forecast | cotir-result | because the abundance of digits is a bit confusing.
KOTIR_D - quotient increment
FORECAST_IN - forecasting inside the sample, i.e. fitting for the whole sample and then forecasting inside - typical forward looking
FORECAST_OUT - forecasting one step ahead outside the sample, i.e. fitting the model on the sample (40 observations) and then forecasting 41 observations. Since this is still on historical data, I can compare it to the fact
RESULT_IN and RESULT_OUT are calculated results. The "-" is the loss.
The whole sample is 5000 samples (read carefully), and I looked at the correlation for the first 500 samples. Trolls, - the calculation is correct. For other variations in length and time interval the ACF will be different, which you and our econometrician have shown. don't worry, keep yourself busy with something useful.
You should also read more carefully what you are asked. You didn't show the whole ACF, just the first 500 samples, although the sample was larger. That's what I was asking about.
Show the whole, all 5000 counts. I'm just wondering what it will look like (according to my research about 90% is an oscillating link model, if you do the right processing of course, although you don't have to...).