Machine learning in trading: theory, models, practice and algo-trading - page 3171
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
...
What conclusion can be drawn?
1. You might get lucky and randomly find a working model :)
2. Without reducing the number of false patterns, it is difficult to build a model using only the principle of greed.
3. You need to develop methods to estimate the regularity observed in a quantum segment or sheet.
4. Randomness does not prove that one is successful in machine learning.
5. A logically valid model is required to be successful.
6. Success on the test sample does not always mean success on the exam sample, and vice versa.
What other conclusions can be drawn?
I lost sight of your quantisation method in the course of the thread, sorry, please remind me.
quantisation of time series is very important, for example, quantising a price by renko method we will get SB (at least statistically relevant), so the expression "quantise it, don't quantise it, you will still get xxx" seems at least understated, because if there is quantisation that destroys information, then there is probably quantisation that extracts information. provided that there is information in tick stream, of course (we fervently hope for it, because application of MO would be senseless otherwise).
Andrey Dik #:
quantising the price using the renko method, we get the SB (at least statistically relevant)
I think no more SB than just bars.
I don't think any more SB than just bars.
I think a much bigger SB, as the stats on bars are very different from SB performance
I don't think anyone here has explored tiki with MO yet.
out loud maybe not, at least not quantising the series.
It seems promising to act "by the contrary method". i.e. to search not for regularities, but for the states of the price (tick) series (I do not want to use "time series"), which are never achievable and do not occur in history.
This will allow to use boundary conditions for building a strategy favourable for traders.
I overlooked in the course of the thread, sorry, the method of your quantisation, please remind me.
The concept of "quantised cutoff" is a simple one - it is the range of the predictor, which has a numerical value on each line of the sample. Anything within the range becomes a unit.
Methods of partitioning into ranges can be different, and I use both variants built in CatBoost (in bousting often used both to reduce the required RAM and reduce the dimensionality), and some of my own, for example, different numerical sequences.
After the predictor is divided into ranges using the obtained grid in one way or another, each segment is taken in turn and evaluated for the value of information in it.
A shift in the probability of belonging to a class by 5% or more from the average value of the sample is considered as valuable information, as well as the number of signals and their distribution in the sample are taken into account.
If the sample with a binary target, we get two groups of quantum segments, in which the probability of hitting 0 or 1 is shifted accordingly.
We create a new sample, where each quantum segment has its own column - if there is a signal in the range - put "1", if there is not - "0".
The answer to that is yes, it will.
I randomly made a choice of the first quantum segment to exclude the signal (string) 1000 times.
Here are a couple of examples of gifs, how the process went with different random first quantum segments (it can be leaves).
And here are static pictures at the moment of intermediate iteration - different stages of selection and randomisation.
What conclusion can be drawn?
1. You may get lucky and find a working model at random :)
2. Without reducing the number of false patterns, it is difficult to build a model using only the principle of greed.
3. You need to develop methods to estimate the regularity observed in a quantum segment or sheet.
4. Randomness does not prove that one is successful in machine learning.
5. A logically valid model is required to be successful.
6. Success on the test sample does not always mean success on the exam sample, and vice versa.
What other conclusions can be drawn?
A shift in the probability of belonging to a class by 5% or more from the sample mean is considered valuable information, as well as the number of signals and their distribution over the sample.
IMHO, it looks like pi-hacking, which Maxim wrote about recently. Unless some stat-tests are used to determine the significance of allocated quanta, it is definitely him.
I once gave a simple example when the best hour of the week for trading was selected on SB (when it obviously doesn't exist). There were only 5*24=120 variants, but it was quite enough that such an hour was always found (the time interval was half a year, I think). There is "sampling stability" there as well.
It seems promising to act "by the contrary method". i.e. to search not for regularities, but for the states of the price (tick) series (I do not want to use "time series"), which are never achievable and do not occur in history.
This will allow using the boundary conditions for building a strategy favourable for traders.
and do the same tests/fits on that series.
Thanks, I'll try MathRand increments.
If you see the same effect (directional OOS dump) - it is the effect of fitting/retraining your TS/MO.
I think, by the definition of the SB there should not be such a situation.