Econometrics: one step ahead forecast - page 88
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
There is no point in analysing the indicator's residuals, because it does not predict anything on its own.
Well, if we assume that the indicator predicts even one bar ahead. What should its residue show then? And why should it show it? The question is probably more for Slava, because what faa says I still have trouble understanding.
Ok, let's assume that the indicator predicts even one bar ahead. What , then, must its residue show? And why should it show it? The question is probably more for Slava, because I still have a hard time understanding what faa says.
It will show that the test results can be trusted.
Vasily, imagine trading without stops with one take profit option = 30 pips. He has done a test, 400 trades and all on take profit - MO=30 pips. But equity drawdowns (which are also residuals/forecasting errors) will be non-stationary, like the quote itself. It means their value will not be described by a normal distribution and the probability of getting a loss exceeding a certain value will not be as small as for HP. There will be "fat tails" in the distribution of drawdowns and it is quite possible that one will get a drawdown and this probability cannot be obtained from tests. So all the other test indicators cannot be trusted either and the MO of this system is not +30 points at all, it's God knows what. This is an exaggerated example :)
In reality, even if the error distribution is stationary, which is quite simple, it does not mean that the drawdown in a series of trades will be the same. The robustness of the system and implies that some properties of the system will remain the same over time - including MO and drawdown in a series of trades. Many evaluate it by eye, by how equity changes over time. Everyone wants a smooth equity - preferably a straight upward angle :)
It will show that the test results can be trusted.
Vasily, imagine trading without stops with one take profit option = 30 pips. He has done a test, 400 trades and all on take profit - MO=30 pips. But equity drawdowns (which are also residuals/forecasting errors) will be non-stationary, like the quote itself. It means their value will not be described by a normal distribution and the probability of getting a loss exceeding a certain value will not be as small as for HP. There will be "fat tails" in the drawdown distribution and it is quite possible to get a drawdown, and this drawdown probability cannot be obtained from tests.
Avals:
C-4:
Why build a vegetable garden and still analyse the residuals from the system? What does it do?
It will show that the test results can be trusted.
Vasily, imagine trading without stops with one take profit option = 30 pips. The person has done a test, 400 trades and all on take profit - MO=30 pips. But equity drawdowns (which are also residuals/forecasting errors) will be non-stationary, like the quote itself. It means their value will not be described by a normal distribution and the probability of getting a loss exceeding a certain value will not be as small as for HP. There will be "fat tails" in the distribution of drawdowns and it is quite possible that one will get a drawdown and this probability cannot be obtained from tests. So all the other test indicators cannot be trusted either and the MO of this system is not +30 points at all, it's God knows what. This is an exaggerated example :)
In reality, even if the error distribution is stationary, which is quite simple, it does not mean that the drawdown in a series of trades will be the same. The robustness of the system and implies that some properties of the system will remain the same over time - including MO and drawdown in a series of trades. Many evaluate it by eye, by how equity changes over time. Everyone wants smooth equity - preferably a straight upward angle :)
It is possible in principle. You have to do the calculation by time intervals. Not count the results of trades, but the results of e.g. days or smaller periods.
It is possible to estimate, but you cannot trust this estimate. One sample will have one statistic, while another will have another and will not converge to the statistic of the whole sample.
There is a sacred cow in tading - the tester.
Using the tester is presented as a revelation. And the adept of this sacred cow, who visits this thread, at the level of an enlightened one, teaches - one should also do a forward test. Of course, discussing the sacred cow is sacrilegious. But I'll give it a try. So, we run the TS through the tester and get its score and see that it is fine. Then, as the enlightened ones teach us, do another run as a forward test and get another result confirming the first one. I write in detail so that everyone can see that exactly two figures are obtained - is that a statistic? My slippers know the answer, but the authors of books with worldwide distribution pass off two figures as statistics and in a non-stationary market!
Long live the sacred cows!
It will show that the test results can be trusted.
...
I.e. the stationary residual would sort of say: "Look, my returnings are normally distributed, which means my variances (spreads) are finite and predictable. I give you 99.8% of my cog that I won't undercut in any transaction by more than three sigmas from my S.C.O.!!!".
And the unsteady residual kind of calls into question the whole calculation part of the model, because it can screw up in such a way that it will override all model calculations: "Well, yes, I have some S.Q.O., - but that doesn't mean anything, because it is the average temperature in my room. On average I can go under 300 points (sko=300 points), but sometimes 1,000 or even 1,500 happens, and it happens much more often than you might expect from a normal distribution."
However, it is not quite clear how you can easily get a stationary residual. In the example above, for example, the residual is our floating eqivity, hence the residual is the movement of the price series, which itself is not stationary, and hence the residual will have all the properties of the price, including non-stationarity. Apparently, faa has already applied his model recursively to the residual:
"Yeah, we got the residual from the model and it's non-stationary. " thinks faa. - "So now we build a model on the residual of the model. Built it. Looked at. The residual is non-stationary again. Repeat the procedure and build the model on the remainder of the model built on the remainder of the original model..." And so on to infinity until the remainder somehow magically becomes stationary. That is, we descend deeper and deeper to an increasingly lower/noisier market level and don't stop until the residual of that noise becomes stationary. But isn't the market fractal? That is, no matter how low the price activity we have to descend to, this activity will still be in the image of large price series: i.e. it will have thick tails and non-stationarity.
Apparently faa applies its model recursively to the residual already:
However, it is not quite clear how we can easily obtain a stationary residual. In the example above, for example, the residual is our floating eqivity, hence the residual is the movement of the price series, which itself is not stationary, and hence the residual will have all the properties of the price, including non-stationarity
"Yep, got the residual from the model and it is non-stationary. " thinks faa. - "So now we build a model on the residual of the model. Built it. Looked at. The residual is non-stationary again. Repeat the procedure and build the model on the remainder of the model built on the remainder of the original model..." And so on to infinity until the remainder somehow magically becomes stationary. That is, we descend deeper and deeper to an increasingly lower/noisier market level and don't stop until the residual of that noise becomes stationary. But isn't the market fractal? I.e., no matter how low the level of price activity we have to descend to, this activity will still follow the pattern of large price series: i.e., have thick tails and non-stationarity.
Absolutely, with a slight addition.
1. There is no point in dealing with a residual of less than a pip.
2. An unsteady residual, i.e. its unsteadiness can be modelled by ARCH.
I don't think this completely solves the problem, but still bites off a piece of it. And if you compare that with the majority modelling of the original kotir, there is a lot of progress compared to them.
That is, the stationary residual says: "Look, my returnees are normally distributed, which means that my variances (scatterplots) are finite and predictable. I give you 99.8% of my cog that I won't undercut in any transaction by more than three sigmas from my S.C.O.!!!".
And the unsteady residual kind of calls into question the whole calculation part of the model, because it can screw up in such a way that it will override all model calculations: "Well, yes, I have some S.Q.O., - but that doesn't mean anything, because it is the average temperature in my room. On average, I can hit 300 points (sko=300 points), but sometimes 1,000 or even 1,500 happens, and it happens far more often than you might expect from a normal distribution."
yes
True, it's not quite clear how you can easily get a stationary residual. In the example above, for example, the residual is our floating eqivity, so the residual is the movement of price series, which itself is not stationary, and therefore the residual will have all the properties of price, including non-stationarity. Apparently, faa recursively applies its model to the residual:
"Yep, got a residual from the model and it's unsteady. " thinks faa. - "So now we build a model on the model residual. Built it. Looked at it. The residual is non-stationary again. Repeat the procedure and build the model on the remainder of the model built on the remainder of the original model..." And so on to infinity until the remainder somehow magically becomes stationary. That is, we descend deeper and deeper to an increasingly lower/noisier market level and don't stop until the residual of that noise becomes stationary. But isn't the market fractal? I.e. no matter how low the level of price activity we have to descend to, that activity will still be in the image of large price series: i.e. have thick tails and non-stationarity.
Yes, in this way it is more likely to get not stationarity, but only a fit to it.