Machine learning in trading: theory, models, practice and algo-trading - page 3061

[Aleksey Vyazmikin]

Maxim Dmitrievsky #:

Predictors and tritment are different. Causal works with outcomes, to determine ATE (average treatment effect) if no traits are involved, or CATE (conditional ... ...) if there are covariates (traits), taking into account some external influence (e.g. an advertising campaign, which is the tritment). This is to determine if there was an effect on the control group. Then, given this effect, you can analyse and improve your model or some metric.

... you're the one who's trying to figure out how to apply something you don't know how to apply.

take a free little introductory course.

I've read two articles by these - they are the ones that gave me the ideas in conjunction with the video.

https://habr.com/ru/companies/ru_mts/articles/485980/

https://habr.com/ru/companies/ru_mts/articles/485976/ 

Well, I'm working on a similar task - that's why I have my own view of nuances. But yes - the new terminology is confusing.

Subtracting probabilities of model predictions - well, I don't think it's any good, especially if you know how these probabilities are distributed in CB.

If you will experiment with libraries specialising on these issues - share your results and your view of the situation.

As long as the main idea of their approach slips in - the main thing is experience and knowledge of the analyst about the subject of analysis :)

[Maxim Dmitrievsky]

Aleksey Vyazmikin #:

I've read two articles by these - they were the basis for the ideas in conjunction with the video.

https://habr.com/ru/companies/ru_mts/articles/485980/

https://habr.com/ru/companies/ru_mts/articles/485976/ 

Well, I'm working on a similar task, so I have my own view of the nuances. But yes, the new terminology is confusing.

Subtracting probabilities of model predictions - well, I don't think it's a good idea, especially if you know how these probabilities are distributed in CB.

If you will experiment with libraries specialising on these issues - share your results and vision of the situation.

As long as the main idea of their approach slips in - the main thing is experience and knowledge of the analyst about the subject of analysis :)

toolkit is offered, work it out.

if you think too abstractly, everyone is working on a "similar task" (only they can't formulate which one).

[Aleksey Vyazmikin]

Maxim Dmitrievsky #:

the toolkit's on offer, figure it out.

If you think too abstractly, everyone is working on a "similar task" (only they can't formulate which one).

I have a separate thread on the forum to solve the problem. The goal is just to establish whether a new factor influencing the probability distribution of the predictor (quantum segment) will appear or not.

Although I haven't worked on it for a long time. More precisely, I need to transfer ideas from paper to code.

It's hard for me to deal with such a thing without basic knowledge of python or er. I have no free time at all and get tired quickly. Although pills have helped a bit - but I take them when I feel like it....

[Maxim Dmitrievsky]

Aleksey Vyazmikin #:

I have a separate thread on the forum on solving the problem. The goal is to establish whether a new factor influencing the probability distribution of the predictor (quantum segment) will appear or not.

Although I haven't worked on it for a long time. More precisely, I need to transfer ideas from paper to code.

It's hard for me to deal with such a thing without basic knowledge of python or er. I don't have any free time at all and I get tired quickly. Although pills have helped a bit - but I take them when I feel like it....

I have to set tasks that I can do. There's no way to read what it says. The frequency has a timeline... too blatant. Then my finger reflexively poked the cross ❌

[Aleksey Vyazmikin]

Maxim Dmitrievsky #:
You have to set your own goals. There's no way to read what it says. The frequency has a time scale... too blatant. Then my finger reflexively poked the cross ❌

Yes, it was the time scale that I took, it's the only way to standardise the independent measures in that sample, and the application value is not lost.

I agree that the problem is difficult and I may not find the solution. However, I see that this is the main reason why models stop working - the probability shift distribution in the predictor range changes a lot with time. And here we can either look for the reason why this happens - to detect the moment of appearance of a new factor, or to look for preconditions for variability from the history of "life" - a kind of survival task.

If we look at your approach through this knowledge, then you indirectly look for areas in the training period where the distributions are stable for the predictors that are significant at that moment, screening out other areas with a different distribution according to some criteria. However, the fact that these areas in the sample are different for different predictors causes a very large sampling period to be cut off. Try to reduce the number of predictors at each iteration - this will reduce the conflict of probability bias between predictors, and thus can increase Recall.

[Maxim Dmitrievsky]

Frequency doesn't have a timeline. I'm just working with model errors :) this approach voiced a long time ago, didn't know about the uplift technique. Turned out to have done about the same thing. What's the point of looking at these distributions? Visualisation of what is already clear.

[Aleksey Vyazmikin]

Maxim Dmitrievsky #:
Frequency doesn't have a timeline. I'm just working with model errors :) this approach voiced a long time ago, didn't know about the uplift technique. Turned out to have done about the same thing. What's the point of looking at these distributions? Visualisation of something that is already clear.

Maybe I'm wrong about the term, what else do you call the frequency of an event over an allotted moment in time?

I'm not talking about visualisation... I'm talking about how to deal with this problem more effectively.

[Maxim Dmitrievsky]

Aleksey Vyazmikin #:

Maybe I'm wrong with the term, what else do you call the frequency of an event over an allotted moment in time?

I'm not talking about visualisation... I'm talking about how to work more effectively with this problem.

It's all there in the thread, you can google the rest. I threw in a book on causal.

Exactly, because of a lot of terms not in the topic, the meaning of what is happening slips away. Although the task itself may be very simple.

[Vladimir Perervenko]

I don't know what attracted you to this topic. For R users, here are a number of packages on this topic. Might help to understand or use. R 4.2.3/4.3.0

other attached packages:
 [1] regmedint_1.0.0               PSweight_1.1.8                MatchIt_4.5.3                
 [4] InvariantCausalPrediction_0.8 mboost_2.9-7                  stabs_0.6-4                  
 [7] glmnet_4.1-7                  Matrix_1.5-4                  grangers_0.1.0               
[10] dagitty_0.3-1                 CompareCausalNetworks_0.2.6.2 CERFIT_0.1.0                 
[13] causalweight_1.0.4            ranger_0.15.1                 causalsens_0.1.2             
[16] CausalQueries_0.1.0           Rcpp_1.0.10                   dplyr_1.1.2                  
[19] causalPAF_1.2.5               causaloptim_0.9.7             igraph_1.4.2                 
[22] CausalMBSTS_0.1.1             KFAS_1.5.0                    CausalKinetiX_0.2.1          
[25] CausalImpact_1.3.0            bsts_0.9.9                    xts_0.13.1                   
[28] zoo_1.8-12                    BoomSpikeSlab_1.2.5           Boom_0.9.11                  
[31] CausalGAM_0.1-4               gam_1.22-2                    foreach_1.5.2                
[34] causaleffect_1.3.15           causaldrf_0.4.2               causaldata_0.1.3             
[37] causalCmprsk_1.1.0            causal.decomp_0.1.0       

Only applying "sr" so far.

Good luck

[Maxim Dmitrievsky]

The packet lord is here