Machine learning in trading: theory, models, practice and algo-trading - page 323
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
Well, why? I've seen publications for EURUSD on M1.
Look at rugarch.
There are a lot of these GARCNs. They have three groups of parameters: the model itself, the type of average, and the type of the residual distribution. For each of the parameter types, the latest peeps. Detrending is discussed above. In GARCH, detrending with ARFIMA, i.e. with fractional differentiation (Hurst).
I'm in the middle of it right now.
Autocorr function M1. The window is 60m.
She's great). At +/-1m there is already zero, or rather a very weak negative. The recommendation of the video, however, is to do the differentiation and then... In our case, after differentiation there is nothing but noise.
I'm working on it right now.
Autocorr function M1. 60m window.
She is great). At +/-1m it is already zero, or rather a very weak negative. The recommendation of the video, however, is to do the differentiation and then... In our case, after differentiation there is nothing but noise.
What if it is 6000 m? Noise is not noise, there should be mini cycles, in theory, one day they will be periodic, and one day they will not
and if 6000 m?
All the same. Delta function, though).
If there is correlation in the direction of the previous readings, it should come out. But it doesn't. It doesn't. If there are cycles, there should be a significant correlation, because several neighboring samples are interdependent in this case, and the peak should expand, even if the cycles themselves are not detected.
SZW The window is sliding, i.e. entire sample is ~52000 samples.
All the same. Delta function, though).
If there is correlation in the direction of the previous readings, it should come out. But it doesn't. It doesn't. If there are cycles, there should also be meaningful correlation, because several neighboring samples, in this case, are interdependent.
SZY The window is sliding, i.e. the entire sample is ~52000 counts.
sad :)
What if we use rsi autocorrect? Or a smoother oscillator. rsi, by the way, does not really depend on the slope of the trend - I changed the slope of the graphs and it showed approximately the same as in the original
And also, as an option, I wanted to try this one https://www.mql5.com/ru/articles/1472
It looks cyclic. you can shove it right into ns or try it with autocorrelation. And its prediction ability is better than rsi's, in my opinion. And it's already multicurrency, by the way, i.e. it depends on a basket of currency pairs, not on the current one.
The only thing I need is to rewrite it on MT5.
sad :)
How about using an autocorrelation psi or a smoother oscillator? Rci, by the way, does not really depend on the slope of the trend - I changed the slope of the charts and it showed approximately the same as in the original
Already considered something like this. Autocor function will reflect only the period of the RSI itself. We calculate on the MAH - there will be a period of MAH smoothing, and so on. Which is natural. In other words, it will have nothing to do with the market.
To be honest, I do not see anything in clusters that is fundamentally different from the same MAH. Imho, of course, but the same eggs in profile.
I'm working on it right now.
Autocorr function M1. 60m window.
She is great). At +/-1m it is already zero, or rather a very weak negative. The recommendation of the video, however, is to do the differentiation and then... In our case, after differentiation there is nothing but noise.
No, it doesn't work that way.
You have to look at the kotier and pick up the tools to solve the problems you have found.
The initial quotient is NOT stationary - a variable average. Lags + deviations from the trend can easily drain the deposit.
There are two ways:
2. Remove trend: average = const
3. Look at the result. More precisely, look at the residual.
3.1 If the residual is stationary, then the ARMA model. There are such series, but very rarely.
3.2 If the residual is NOT stationary, then differentiate again. ARIMA model. Rows for this model are more frequent, but still very rare.
4. Looking at the residual and modeling GARCH.
In reality it is much more complicated
Already considered something like this. Autocor function will reflect only the period of the RSI itself. We calculate by MASK - there will be a period of MASK smoothing, etc. Which is natural. In other words, it will have nothing to do with the market.
To be honest, I do not see anything in clusters that is fundamentally different from the same MAH. Of course, this is imho, but it's the same as eggs in profile.
Well, and the last option - to train the spreader, thanks metaquotes soon promise custom feeds, where you can build all sorts of tools using standard tools
on the same stocks or indices
No, it doesn't work that way.
We need to look at the quotient and select tools to solve the identified problems.
1. the initial quotient is NOT stationary - a variable average. we can trade trends, but we cannot distinguish between a correction and a reversal. A lag + deviations from the trend can easily drain the deposit.
It is understandable. However, if we start to apply all mentioned above to a Wiener process (random walks), then we see trends, reversals, flats, and the hell with it - it has been tried before). We will calculate all sorts of regressions. (It would only be useless.) And, as the same Wiener or Feynman wrote, before solving a problem, it is a good idea to find out if it has a solution.
To do this, first of all, it is necessary to find any stable correlation relations (their existence), and then build models. That is how it seems to me.
However, so far there is silence.
I recently did an experiment like this. At each point of the time series I built a polynomial regression for the previous period, and displayed only the last point. Calculation is long, about 8 hours - did not save anything, and can not show. Only in words. I'll probably show it later, on a piece.
So, cyclically, there is a kink in the regression line, after which there is a smooth line again. I have to say that I haven't figured out why this happens, but we can assume that in the vicinity of these points, the statistics of the time series change in leaps and bounds.
PS found a piece of the graph.
Ignore the outliers (don't know where they come from, maybe the polynomial coefficients are off the charts). Unfortunately I can not combine this particular graph with the price series.
Lately I did an experiment like this. At each point in the time series I built a polynomial regression for the previous period, and displayed only the last point. Calculation is long, about 8 hours - did not save anything, and can not show. Only in words. I'll probably show it later, on a piece.
So, cyclically there is a kink in the regression line, after which there is a smooth line again. I must say that I understand why it happens, but I can assume that in the vicinity of these points, the statistics of time series changes abruptly.
What does it mean when during genetic optimization the results start to go into turbulence? :) The graph should improve with time