Machine learning in trading: theory, models, practice and algo-trading - page 2741

[JeeyCi]

mytarmailS #:

In a sliding window retrain the model and look at the importance of the features, or just take some good feature identifier and look at it in the sk. Window.

...

Also different feature-selectors to suit all tastes, probably 5% of what is available in R-ka.

floating window size to get effective estimations, not just floating by step or const , - that's the problem - at each iteration to adjust the window size only - the model will take a long time to learn ... and periodic manual retraining is itself a sliding window! -- you will do it periodically anyway (going beyond acceptable st.dev) -- if you have your own retraining schedule - you can automate it too. BUT I repeat - the window size is also floating.

...

different? - it still algorithmically comes down to feature mapping always (!), whatever you want to call it, ... just its own nuances and its own field of application.

even if not everyone wants to call it correlation

SanSanych Fomenko # :

I'm relying on the notion of predictor-teacher correlation. "Linkage" is NOT correlation or the "importance" of predictors from fitting almost any MOE model.

caret link - same Classification and Regression Training - as a trivial MO, same sklearn for Python.

It's just that MO was created not only for building probabilistic models (based on existing probability distributions), but also for determinantal and dynamical ones... but the basis of any generalisation of probabilities will always (!) be statistics (with its correlations), whatever you call it... Otherwise you will get biased(!) estimates - i.e. your model will be modelling something else (at random), not your target.

[Renat Akhtyamov]

mytarmailS #:
What do we need to do to stop fighting and unite for one goal????????

great question!

---

we need to recognise whether there are more buyers or sellers.

but the only trouble is, if the price goes down, there will be more buyers, guaranteed!

[СанСаныч Фоменко]

JeeyCi #:

caret link - same Classification and Regression Training - as a trivial MO, same sklearn for Python.

Read more carefully, and you don't need to make stuff up

[СанСаныч Фоменко]

mytarmailS #:
The script calculates the importance of predictors in a sliding window using two different algorithms, foremost and one other way... Just like you asked.

Forrest gives the frequency of predictor use in a particular algorithm, so it easily gives high importance to predictors that are NOT related to the target.

[JeeyCi]

СанСаныч Фоменко #:

Read more carefully and don't make stuff up.

You can't read more carefully if caret is actually spelt that way.

[mytarmailS]

СанСаныч Фоменко #:

Forrest give the frequency of predictor use in a particular algorithm, so it is easy to give high importance to predictors that are NOT related to the target.

I repeat: there are dozens of ready-made fiche selectors, and everyone is looking for a connection with the target, everything has already been invented before us, long ago.

[Aleksey Nikolayev]

СанСаныч Фоменко #:

I think I've seen some hints on the application of survival analysis. Is there anything interesting in this direction? I have some ideas related to replacing the breakdown time with the value of maximum price movement in the desired direction during a trade before the stop is triggered. The basic idea is to look for deviations from what the behaviour should be for random wandering. In this area, by the way, the application of matstat (Cox regression, for example) and MO is also very developed.

[JeeyCi]

JeeyCi #:

... window size is also a floating window size

ML falls short of a suitable tool precisely because

ML faces distinct challenges in the RL context where the data is generated by the interaction of the model with the environment using a (possibly randomised) policy

i.e. first you have to model the current environment, and then the fs in it and the corresponding behaviour there, i.e. the actor must have a state -- this is the basis for switching to a new window (aka taking into account the floating window size) and thus to a new environment policy and the corresponding/new behaviour of the actor in it ... in general, Deep Reinforcement Learning is probably more suitable, in which

To train a trading agent, we need to create a market environment that provides price and other information, offers trading-related actions, and keeps track of the portfolio to reward the agent accordingly.

example - there is game theory (actor'a interaction with the environment) and information theory (tangency of information to actor'a responses and environment reactions, i.e. new environment conditions being formed, aka consequences)... - I, by the way, did not understand what SanSanych Fomenko meant by the term information theory . .. or is his thesis being misread again?

p.s. I haven't tested the example myself....

[СанСаныч Фоменко]

JeeyCi #:

I can't be more careful if caret is really what it stands for.

It says classDist {caret}, i.e. it specifies a particular function that is part of the caret PACKAGE

As I understand, you don't know R. Then why are you wasting your time on this thread, and on MO in general?

Without mastery of R, discussion of MO is meaningless.

[СанСаныч Фоменко]

Aleksey Nikolayev #:

I think I saw some hints of a survival analysis application.

I haven't done it myself, but I once asked a similar question. This approach seems promising to me, but it is from a completely different sphere.