Machine learning in trading: theory, models, practice and algo-trading - page 1189

[Maxim Dmitrievsky]

Aleksey Nikolayev:

I haven't seen anything in the thread about information criteria(Bayesian or Akaike). Perhaps they are used by default (in the applied MO packages)?

cross entropy is used or log-loss

for multiclass and binary classification

For regression tasks, mean-squared. (rms) and similar

I guess this is Akaike

[Maxim Dmitrievsky]

checked PCA and LDA for pre-processing predictors (getting rid of collinearity). As expected, it doesn't work on new data because the components themselves start to bounce, although the model may be a little better trained on them in the train. But because components themselves behave unpredictably on new data, it turns out the same and even worse. In general, many of the classical MO techniques are simply not applicable to the market, or rather it does not work directly.

[Aleksey Nikolayev]

Maxim Dmitrievsky:

Cross entropy is used or Log Loss

for multiclass and binary classification respectively

For regression problems, the rms. (rms) and similar

I take it this is the Akaike.

Looks like it is.

[Aleksey Nikolayev]

Maxim Dmitrievsky:
Checked PCA and LDA for pre-processing predictors (getting rid of collinearity). As expected it doesn't work with the new data, because the components themselves start to jump, though the model may be trained on them a little bit better in the train. But because components themselves behave unpredictably on new data, it turns out the same and even worse. In general, many of the classical MO techniques are simply not applicable to the market.

Because of non-stationarity, it is sometimes necessary to discard obsolete history:

1) You need the right algorithm for discarding the obsolete part of the sample (search for discontinuity).

2) The remaining part of the sample will always be of variable length and often short. We need to get the right models for this too.

[Igor Makanu]

Maxim Dmitrievsky:

This is already implemented in the new article, but not exactly as we would like.

What article are we talking about?

[Maxim Dmitrievsky]

Igor Makanu:

which article are we talking about?

which has been submitted for moderation, has not yet been published

[Aleksey Nikolayev]

Maxim Dmitrievsky:

There is another approach, I see it as the most promising at the moment - bruteforcing models through enumeration of output variables

Roughly speaking, the virtual trader trades a pseudo-random number of times (like Monte Carlo or genetics), each time he/she watches his/her trade and corrects wrong positions, roughly speaking flips loss-making ones to make them profitable

after each pass it retrains on corrected trades. This is already implemented in the new article, but not quite as we would like it. We need interesting combinations of exits that depend on current market conditions, such as dispersion and slope over n-bars. For each such characteristic a distribution is selected from which random trades are sampled, then in the same way loss trades are corrected and trained on them. Through a lot of passes we look for the optimal strategy (based on minimal errors on the test sample).

Dear Expert Advisors, pay attention, Question: How to organize in an interesting way the dependencies of the current probabilistic characteristics of the market and distributions from which random outputs are sampled? In this case both the number of trades and some dependencies inside the model will change, i.e. we obtain a lot of various profitable models (solutions), among which the optimal model is selected according to a custom optimization criterion (model error, stability on new data).

The exits seem sensible if the capital drawdown is higher than the specified one and its growth is too slow compared to the volatility (small ratio of drift to variance)

It is not very clear how this approach will help to fight non-stationarity.)

[Igor Makanu]

Maxim Dmitrievsky:

which has been submitted for moderation, hasn't been published yet.

Thanks, I will wait, now I will not miss the article!

[mytarmailS]

interesting articlehttp://www.long-short.pro/post/indikator-haosa-i-rezhimy-fondovogo-rynka-886

and the blog in general

[Aleksey Vyazmikin]

mytarmailS:

interesting articlehttp://www.long-short.pro/post/indikator-haosa-i-rezhimy-fondovogo-rynka-886

And the blog in general.

Have you tried expressing his indicator in code?

1) находим максимум минус минимум для каждого из 10 дней, предшествующих настоящему моменту; 
2) берем сумму этих значений (сумма частей); 
3) находим 10-дневный диапазон: 10-дневный максимум минус 10-дневный минимум; 
4) делим сумму частей на целый диапазон – это основная мера фрактальной размерности/сложности; 
5) берем 60-дневную среднюю 10-дневной серии значений сложности – это метрика квартального хаоса/стабильности; 
6) используем 252-дневное нормальное распределение z-оценки или ранг процентиля метрики хаоса/стабильности; 
7) значения, которые выше 0,5, показывают, что рынок находится в режиме «хаоса» и гораздо менее предсказуем и нестационарен, значения ниже 0,5 показывают, что рынок стабилен и намного более предсказуем.