Machine learning in trading: theory, models, practice and algo-trading - page 2752
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
Correct. For lack of a priori assumptions, the second type is used. I wonder how Sanych sees it.
As I see it, an a priori assumption is made that each class is given by a Gaussian distribution, which gradually changes over time, due to non-stationarity. Without such an assumption, using the Mahalanobis distance approach makes little sense.
Personally, I find such an assumption too strong to be true for any instrument at any time interval.
Just to be on the safe side, I am not going to tell you what you should do. Rather, I am just thinking aloud what I would do myself in case there is an algorithm that gives good results but is poorly implemented in the form of an Expert Advisor (although I always try to avoid this option). Most likely, I would try to get similar results using algorithms that are easier to implement. Among other things, I would get some analysis of what exactly the algorithm is good at.
I would start with KNN and if it gives a similar result, then it is a matter of good selection of a common set of predictors. If the result is much worse, then maybe it's just a matter of choosing a subset of predictors at each moment. To test this hypothesis, I would try to use local regression (something like LOESS), since regression already allows you to compare the significance of predictors. Further steps are already based on the results of the analysis. By the way, with the appearance of matrices in mql5, linear regression has become easy to implement directly in it.
It is funny to say, but any variants of matrix-smoothing do NOT have predictive ability, including LOESS, I have tried it.
It's funny to say, but any variants of mashek-smoothing do NOT have predictive power , including LOESS, tried it.
Not funny or informative in general. also gave up averaging, but I realise that the range of my studies is small, not formalised and not specific. You can't get specificity in old studies, but at least they could describe them in a formalised way for understanding. What averages are investigated, in what range and on what periods.
ZY my data 18-20 years, mash 3, 14, 60, 120 on all TF fours, some BB. The best result on TF 1 hour, even spread drain. Selection of parameters manually.
It's funny to say, but any variants of mashek-smoothing do NOT have predictive power , including LOESS, tried it.
MAs by nature do not have predictive power. They are simply not about that, even if they are matrices (what exactly is wanted here is to multiply a vector of prices by an N-dimensional matrix to get a profit; the whole fuss is about matrix preparation and correction).
They are about "what price to consider valid at T time ago", which is generally speaking extremely important. What was T time ago, from the height of what has already been lived. They are about more or less interpretation (understanding) of history.
used classDist
1) training/retraining the model when classDist is about 1 , that is with filtering on good states from the point of view of the algorithm
2) feed classDist as a feature for the whole sample
3) filtering on different states
4) tried simple training and constant retraining
in all cases the predictive power is not better than stochastic, it is almost random....
So there are questions about the reality of the statements
MAs by nature have no predictive ability. They are simply not about that, even if they are matrices (what exactly is wanted here - to multiply a price vector by an N-dimensional matrix in order to get a profit; the whole fuss is about matrix preparation and correction).
They are about "what price to consider valid at T time ago", which is generally speaking extremely important. What was T time ago, from the height of what has already been lived. They are about more or less interpretation (understanding) of history.
According to the classics, the greatest predictive ability is represented by prices/levels/whatever close to the current one. We try such TS and see frankly weak results. This is not the fault of the mashki.
In reality this is not respected, the current price can be influenced by distant changes and the near ones can be disorientating.
That's why I don't believe in retraining in a sliding window, although sometimes it looks good, like in my entropy article.
It is interesting to identify reference points in history that influenced the future with a delayed effect. This can be done including in a sliding window by original approaches.
If necessary, I can describe it in more detail. In general, it can look like a floating or shifting window, not fixed. Modern algorithms for processing sequences work on approximately the same principle. But for fora will have their own specifics, which they do not take into account.
To determine the specifics, we need to theorise a bit about fractals and their properties. For example, consider a time series as fractal, then there will be something to base on.
Then levels/patterns/something else acquire a real meaning and represent a particular description of a system with known properties. Then it can all be put into one theory and common understanding.
There are various attempts to move in that direction, including fractional differentiation. But something more explicit and stronger is needed.
Econophysics has moved somewhere wrong in my opinion, maybe I don't fully understand it. Lots of formulas and little sense.
think about it
Classically, the greatest predictive ability is represented by prices/levels/whatever close to the current price. We try such TS and see frankly weak results. Mashki is not to blame here.
In reality it is not observed, the current price can be influenced by distant changes, and the nearest ones can disorientate.
That's why I don't believe in retraining in a sliding window, although sometimes it looks good, as in my article about entropy.
It is interesting to identify reference points in history that influenced the future with a delayed effect. This can be done also in a sliding window by original approaches.
If necessary, I can describe it in more detail. In general, it can look like a floating or shifting window, not fixed. Modern algorithms for processing sequences work on approximately the same principle. But for fora will have their own specifics, which they do not take into account.
To determine the specifics, we need to theorise a bit about fractals and their properties. For example, consider a time series as fractal, then there will be something to rely on.
think about it
What will fractality give?
It is interesting to identify reference points in history that influence the future with a delayed effect. This can be done also in a sliding window by original approaches.
And here he, replacing my concept of "Event" with "Repertory Points", will pretend that he was not told about it a dozen days ago.... yeah.
Classically, the greatest predictive ability is represented by prices/levels/whatever close to the current price. We try such TS and see frankly weak results. Mashki is not to blame here.
In reality it is not observed, the current price can be influenced by distant changes, and the nearest ones can disorientate.
That's why I don't believe in retraining in a sliding window, although sometimes it looks good, as in my article about entropy.
It isinteresting to identify reference points in history that influenced the future with a delayed effect. This can be done also in a sliding window by original approaches.
If necessary, I can describe it in more detail. In general, it can look like a floating or shifting window, not fixed. Modern algorithms for processing sequences work on approximately the same principle. But for fora will have their own specifics, which they do not take into account.
To determine the specifics, we need to theorise a bit about fractals and their properties. For example, consider a time series as fractal, then there will be something to rely on.
There are various attempts to move in that direction, including fractional differentiation. But something more explicit and stronger is needed.
Econophysics has moved somewhere wrong, in my opinion, maybe I don't fully understand it. Lots of formulas and little meaning.
think about it
points in history beyond the classical SB are worthy of special attention.
points in history beyond the classical SB are worthy of special attention
including, but it is necessary to determine their nature and somehow describe them in general, in order to use an adequate method including MO
in my opinion, fractal theory is closer to the body in this case