Building a trading system using digital low-pass filters - page 12

 
Prival:


One conclusion that I think is useful from this research is the recommendation of the SL to be set at 4 H for this currency pair, which is 81 points (3*SCO). Who wants, you can download and study your favorite currency pair. If something is unclear about the program and calculations, please contact me on Skype, I will try to help.

This is an interesting conclusion, but it seems to me that you can not use it in practice. The reason is simple, such a value of SL will take into account fluctuations in one-time increments (returns), but stops do not adjust to the trend :) So such a stop with a confidence interval probability would not work for a one-off price change (i.e. 4 hours), I agree with that. But it will easily be picked up by a trend (several H4 bars) developing in the opposite direction to the open position.

Therefore, there is not much utility, except when trading within the 4 hour interval.

What is even sadder is that this estimation is equally valid for TP.
 
bstone:
Prival:


One conclusion that I think is useful from this research is the recommendation of the SL to be set at 4 H for this currency pair, which is 81 points (3*SCO). Who wants, you can download and study your favorite currency pair. If something is unclear about the program and calculations, please contact me on Skype, I will try to help.

This is an interesting conclusion, but it seems to me that it cannot be used in practice. The reason is simple, such a value of SL will take into account fluctuations of one-time increments (returns), but after all stops do not adjust to the trend :) So such a stop with a confidence interval probability would not work for a one-off price change (i.e. 4 hours), I agree with that. But it will easily be picked up by a trend (several H4 bars) developing in the opposite direction to the open position.

Therefore, there is not much utility, except when trading within the 4 hour interval.

What is even sadder is that this estimation is equally valid for TP.

Maybe I didn't put it accurately, 81 it is a level (decision-making level), if it is broken through within 4 hours (in either direction), which means that there is a 95% probability it is not caused by noise. And this knowledge can be used in different ways. Although ...
 
NorthernWind:
Prival:

To NorthernWind

The graphs you presented are not the numerical series that the mathematician is asking for. In 5-10 min. I think I will post studies that confirm the cyclic nature of the candle size.


So, isn't H-L the same thing?

No return is Close(i)-Close(i+1) in MQL. I.e. the value of price increase from bar to bar, not the value of bar H-L. Different sequence of numbers, hence (usually) different characteristics.
 
Prival:
NorthernWind:
Prival:

To NorthernWind

The graphs you presented are not the numerical series that the mathematician is asking for. In 5-10 min. I think I will post studies that confirm the cyclic nature of the candle size.


So, isn't H-L the same thing isn't it?

No return is Close(i)-Close(i+1) in MQL. I.e. the value of price increase from bar to bar, not the value of bar H-L.

That is, we take one and the same series of numbers and draw two samples on it, and the result of the first sample is that the series is non-stationary (at least, we have a non-permanent MO), and the second sample is stationary.
 
No, the original series are different, and the transformation in your version is as non-stationary as it was, but it is difficult to prove mathematically that the return leads to stationarity, but I think you can, otherwise a mathematician will not accept an unsubstantiated and visual proof. I need some help for that, but I think I will be able to do it. But I think I can do it by the weekend.
 
Prival:

Maybe I didn't put it accurately, 81 it's a level (decision-making level) which, if broken within 4 hours (in either direction), indicates that it is 95% likely to be caused by something other than noise. And this knowledge can be used in different ways. Although ...

Well without knowing the characteristics of the noisy signal, there is nothing to do with the knowledge of the differences going beyond the noise :)
 
Prival:
No, the original series are different, and the transformation in your version is as non-stationary as it was, but it is difficult to prove mathematically that the return leads to stationarity, but I think you can, otherwise a mathematician will not accept an unsubstantiated and visual proof. I need some help for that, but I think I will be able to do it. I think I can do it by the weekend.

No, the original series are the same, they are the same price series.
 

to North Wind

Можно не просто МА а набор заранее расчитанных цифровых фильтров. Возможно, что эти самае экстремумы не сильно "гуляют" по шкале.

In addition, these peculiarities of the candlesticks do not give much hope for accuracy, so...

It should be noted, but extremums of the spectrum, if memory serves me correctly, were not very "fluctuating", and I had to check them once a week or so.

to Mathemat

Grasn, on p. 10 there (quoting myself):

It was kind of a joke. There's what it said: :о)

Man, it's a dead end. The definition of stationarity itself... it's not the same, it's not rigid. "For a process to be stationary, the m.o. etc. must be constant". I.e. the m.o. process must itself be stationary :))) It's buttery...

That's not the main point. I'm not much of a mathematician, but say, let's say you found a good way to convert a non-stationary process to a stationary one. So, what next? What will it give you? The ability to predict? Again what, to predict, a stationary process? And why would you want to predict a process that exists (in this case), only in your imagination? How will you then move on to the "real" process?

 

The Hi-Lo distribution must surely be different from the distribution of returns, as it is rather something similar to the distribution of maximum returns. Finally, I won't be too upset if Prival convincingly shows that returns are highly non-stationary.

What's wrong with such an idea of the stationarity test - since famous tests seem to assume a priori some model of the process: take the whole population (say, the same 14 thousand points), calculate its PDF ("global"). Then we take random samples inside the process with sufficient length (say, 1000 points each, and not necessarily going in series, so that we immediately encounter cyclicities, if any), calculate for each a pdf ("sample") and then look at the deviation of the sample p.d.f. from the global in some sense (say, the integral of the square of the difference between pdf and PDF).

After collecting statistics (say, 1000 samples), we build an error distribution, and use it to try to judge how stationary our process is. It seems that this procedure is designed to detect stationarity in the narrow sense (at all points). It seems easy to adapt it to stationarity in a broad sense.

2 grasn:

let's say you've found a good way to convert a non-stationary process to a stationary one. So what next? What does it give you? The possibility to make forecasts? Again what, to predict, a stationary process?

No, my goal is different. I want to create high-quality synthetic histories on my own (based on something stationary). I want to load them into a tester and test the strategy, varying not the system parameters, but the stories. At the same time I will have as much history data, not significantly different from real data by main characteristics as I need, even if it is a billion of samples. And the reliability of testing should grow by an order of magnitude (though what is an order compared to almost zero reliability?).

P.S. Yeah, I screwed up with the stationarity test. I completely forgot about ACF...

 
to Mathemat

No, my goal is different. I want to create high quality synthetic histories myself (based on something stationary). I want to load them into a tester and test a strategy, varying не параметры системы, а истории. At the same time I will have as much historical data, not much different from real ones by main characteristics, as I need, even a billion samples. And the reliability of the test should increase by an order of magnitude (although what is an order of magnitude compared to near-zero reliability?).


Cool. Well, take as a basis a series, which is already stationary by definition (if you search - there are many), why do you need these criteria????

Here, I've come up with one more variant: why not take serial generation of zigzags, where each element y = a + b * x, parameters a, b, N (segment length) you set "randomly". Plus you impose noise. Necessary distributions can be "looked up" from real zigzags. What's wrong with that?

By the way, there are simply ready-made methods of signal generation by distribution of a given kind.

In short, what is the main problem? I really do not really understand how such series will be useful to you for testing the Expert Advisor - but that will be too much for you.