Principles of working with an optimiser and basic ways of avoiding fitting in. - page 3
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
Why does a pattern need stationarity? Suppose we have a working pattern. The distribution of its occurrence over time is rigidly nonnormal. The main characteristics of this pattern are also nonstationary and float with time. So what? The main condition is only one, - that it continues to appear and not disappear. Simply our MO will be non-stationary, but still positive, and this is the main thing. Another matter is that non-stationarity seriously complicates the search for these very patterns. We cannot rely on standard methods of statistics to identify it and in the process use it. For example, if it, appeared every day for the last year, and suddenly disappeared today, statistics will say - the pattern no longer works. But this is not true, because it appears when it pleases, and is not obliged to generate stationary characteristics. This is its property, at a fundamental level, that determines the need to reoptimise algorithms. Because one way or another, we are working with fixed parameters that correspond perfectly to a given pattern only on history. Tomorrow it will be slightly different, which means that there will be a shift from the extremum of our fitting.
And it's just a matter of surviving tomorrow's shift. And we can survive using relatively stable regularities, or (and) sufficiently rough (simple) methods of identification and dealing with them that their rough estimation would allow the regularity itself to change within sufficiently wide limits.
This is my rationale, why simple methods are usually more effective than complex ones, and why it becomes possible to make money in the market in the first place.
I did some digging and I found -
Stationarity is the property of a process not changing its characteristics over time.
Thus, a non-stationary series changes its characteristics over time. But that does not mean that there cannot be a pattern in it.
You are confusing a non-stationary, financial series with a chaotic series. A chaotic series cannot have regularities, but a non-stationary series that changes its characteristics over time can. Moreover, there can be regularities that predetermine these changes.
Even at first glance there are some regularities visible in the financial series -
A clear upward and downward movement in the form of a trend. A pattern? - A pattern.
A pronounced indefinite sideways movement as a flat. Regularity? - Regularity.
Shoulders-heads, flags and other shapes. Regularity? - Regularity.
And much more......)))))
Why does a pattern need stationarity? Suppose we have a working pattern. The distribution of its occurrence over time is rigidly nonnormal. The main characteristics of that pattern are also non-stationary and float with time. So what? The main condition is only one, - that it continues to appear and not disappear. Simply our MO will be non-stationary, but still positive, and this is the main thing. Another matter is that non-stationarity seriously complicates a search of these very regularities. We can't rely on standard statistical methods to identify it and in the process use it. For example, if it, a pattern, has been appearing every day for the last year, and suddenly disappeared today, statistics will say that the pattern no longer works. But that's not true, because it appears when it wants to, and doesn't have to generate stationary characteristics. It is its property, at a fundamental level, that determines the need to reoptimise the algorithms. Because one way or another, we are working with fixed parameters that correspond perfectly to a given pattern only on history. Tomorrow the pattern will be slightly different, which means that there will be a shift from the extremum of our fitting.
This is quasi-stationarity - a change of mo in a certain range. Maybe it's not just about mo, but in this context it's the one we're most interested in.
And the question of all questions is just surviving the shift of tomorrow. And we can survive with relatively stable regularities or (and) sufficiently rough (simple) methods of identification and work with them, so that their rough estimation allows the regularity itself to change within a wide enough range.
This is my rationale, why simple methods are usually more effective than complex ones, and why it becomes possible to make money in the market in the first place.
So there can be a super complex method, but a rather crude estimation of the pattern.) It's more a question of the number of system parameters and the sensitivity of the result to their change. If a small change in the parameter causes a change in the result, this is not good. There are other signs. I just recently wrote about this here https://www.mql5.com/ru/forum/137614/page5
try to take the remainder not always, but selectively in chunks. If you know how to identify the beginning and end of such chunks on a row (not ex post facto, of course), this will be enough to trade. If not, you will have to change the model.
Once again: there are no stationary pieces.
Once again: you want to gain positive values and no more than predefined losses in a trade. These are the quasi-stationary parts from the entry to the exit. And they are of course on the traded price series.
Equity increments are quasi-stationary with positive mo varying preferably within small limits. Otherwise, there is no need for such an equitability and the system generating it.
faa1947: Еще раз: не бывает стационарных кусов.
Once again: entering a trade you want to get positive mo and no more than a predetermined loss? These are the quasi-stationary stretches from trade entry to exit.
That's the point of this thread.
In fact it is quasi-stationary, in prediction it is non-stationary. The test, including the forward test, is quasi-stationary, while the future is non-stationary, and therefore the test tells us nothing. It is necessary to translate the future quotient into a quasi-stationary state. This can only be done by simulating non-stationarity, at least in part.
This can only be done by modelling non-stationarity, at least in part.
Who minds, model it)) But anyway, when modelling market change, you have to rely on its statistics in the past and on some kind of unchanging model. That is, only the parameters of this model change based on the nearest history. Adaptability is a normal property of TS :)