Machine learning in trading: theory, models, practice and algo-trading - page 2605

[mytarmailS]

Maxim Dmitrievsky #:

Rumor has it that tambourine dancing is no longer in vogue

They say that if you have nothing to say, it's better to keep silent.

[Доктор]

Aleksey Nikolayev #:

Even if you are right, it is a problem with the premise, not with the construction. But I suppose there is some catch in your strategies. For example, the profit is not much higher than the interest on deposits, or its volatility is too high. Otherwise, you could hope to find and assemble many such strategies into a larger portfolio.

Profit percentage is triple-digit, volatility is minimal. But it is not scalable, and it is possible to create a portfolio only by increasing the number of instruments (markets).This strategy is not mine.)

[Aleksey Nikolayev]

Doctor #:

Profit percentage is triple-digit, volatility is minimal. But it does not scale, and it is possible to collect a portfolio only by increasing the number of instruments (markets).The strategy is not mine).

I unconditionally believe in everything written there. I only regret that there are no results for the last two years)

[Доктор]

Aleksey Nikolayev #:

I unconditionally believe everything written there. I only regret that there are no results for the last two years)

The author is widely known in narrow circles. When asked "will there be results for the year 20XX" usually answers that there is no point in writing, because from year to year the result is the same. Last entry on the resource dated March 29, 2021.

[Renat Akhtyamov]

Doctor #:

The author is widely known in narrow circles. When asked "will there be results for 20XX year" he usually answers that there is no point in writing, because from year to year the result is the same. The last entry on the resource is dated March 29, 2021

leaked....

[Renat Akhtyamov]

Aleksey Nikolayev #:

1) I think it is obvious that there is not and cannot be any way to prove that a pattern established in history will necessarily work in the future.

2) The existence of a method that establishes a deterministic (non-random) pattern for the future based on data from the past would be a negation of (1)

We only have crossvalidation, which can only establish the homogeneity of a pattern on history. We can only interpolate the pattern, not extrapolate it. We have only a very weak PROPOSAL that a well interpolated pattern will turn out to be well extrapolated. This is not a deductive inference, but only an inductive one - a variant of inference by analogy.

Regularities will emerge if the system is written, not philosophized

Once again I tell you

write the system with the mindset of not sinking, not making money.

---

in order not to be unsubstantiated, you will understand/detect/see the kotir's reactions to certain actions of the system

usually it's:

- GEP

- black swan

- trend

- flat

- hairpins and long tails

All these regularities are in no way connected with the causes generally accepted by theorists

theory, which is necessary and sufficient to know to build a trading systemhere

[Aleksey Nikolayev]

Doctor #:

The author is widely known in narrow circles. When asked "will there be results for 20XX year" he usually answers that there is no point in writing, because from year to year the result is the same. The last entry on the resource is dated March 29, 2021.

Don't get me wrong -- not that I question what this particular person has written or your recommendation given to him. It's just that this forum, as well as many others, is full of statements similar in form to "We have THIS kind of equipment, but we won't tell you about it!" The above is a good example) If we believe in one case, what reason can we have to refuse to believe in others like it? But it is quite impossible to build anything meaningful on this very shaky ground. Therefore, I prefer to base it onassertions, though weaker, but meaningful and at least somehow verifiable.

[mytarmailS]

Aleksey Nikolayev #:

I unconditionally believe everything written there. I only regret that there are no results for the last two years)

Aleksey Nikolayev #:: Aleksey Nikolayev #:: Aleksey Nikolayev.

Aleksey, do you trade in general or in search?

[Aleksey Nikolayev]

mytarmailS #:
ahahahah 5 s ±

Alexey, are you trading at all or are you looking for something?

I'm trying to do both, but so far I'm getting a noticeable bias towards theory.

[Maxim Dmitrievsky]

There is such a formulation of the question:

Two models are used. One predicts to buy or sell, the other to trade or not.

First the first model is trained, then we see where it predicts badly, mark those examples as "not to trade", the other good ones as "to trade", then we train the second model.

The first model is tested not only in the training area but also in the additional one, while the second model is trained in both areas.

We repeat this several times, retraining both models on the same dataset. The results gradually improve on the samples. But not always on the control sample.

Parallel to this there is a log of bad trades cumulative for all passes, all "bad" deals for "not to trade" are collected in it for training the second model and filtered by some principle, like the more copies of bad deals for all passes, the more chance to mark them as "not to trade"

to_mark = BAD_SAMPLES_BOOK.value_counts()
mean = to_mark.mean()
marked_idx = to_mark[to_mark > mean*bad_samples_fraction].index
pr2.loc[pr2.index.isin(marked_idx), 'meta_labels'] = 0.0

For example, for each date some amount of bad trades is accumulated for all iterations of training, where this number exceeds the threshold (mean, average), those trades are marked as "do not trade". The rest trades are skipped, otherwise it would be possible to exclude all trades if there are a lot of training iterations.

bad_samples_fraction

coefficient allows to adjust the number of exit trades, the lower it is the more trades are filtered out

... by this point i'm already tired of writing ...

How can such a combination of models be improved so that it improves its results on a new independent plot?
Is there any philosophy as to why this might work? Other than the fact that the models naturally improve each other (fall errors) on each round of retraining, but how do you get rid of the fit?

Illustration. The graph is broken into 3 parts. The last one trains the first model, the penultimate and last are the second, the first third is an exam sample. Naturally the last part will be the best and the first third the worst.

Here we had 15 iterations of retraining both models, using the log of bad trades.

Iteration: 0, R^2: 0.025863859193577587
Iteration: 1, R^2: 0.20881945768090338
Iteration: 2, R^2: 0.38691567117849557
Iteration: 3, R^2: 0.8538667616323108
Iteration: 4, R^2: 0.6289257079331403
Iteration: 5, R^2: 0.49590724745042913
Iteration: 6, R^2: 0.6899198178561211
Iteration: 7, R^2: 0.7914478307518835
Iteration: 8, R^2: 0.6271633947453318
Iteration: 9, R^2: 0.5022724259087565
Iteration: 10, R^2: 0.8568310685006555
Iteration: 11, R^2: 0.042448644454852524
Iteration: 12, R^2: -0.17980715185584073
Iteration: 13, R^2: 0.8294648122002825
Iteration: 14, R^2: 0.7615234602466088