Machine learning in trading: theory, models, practice and algo-trading - page 3344

New comment
 
Aleksey Nikolayev #:

We need a good probabilistic forecasting for series, but not as cheesy as it is nowadays (quantile regression, for example). I didn't see it in the article itself, though the list of literature seems to contain it.

There is something from Yandex

Uncertainty in Gradient Boosting via Ensembles
Uncertainty in Gradient Boosting via Ensembles
  • research.yandex.com
For many practical, high-risk applications, it is essential to quantify uncertainty in a model's predictions to avoid costly mistakes. While predictive uncertainty is widely studied for neural networks, the topic seems to be under-explored for models based on gradient boosting. However, gradient boosting often achieves state-of-the-art results on tabular data. This work examines a probabilistic ensemble-based framework for deriving uncertainty estimates in the predictions of gradient boosting classification and regression models. We conducted experiments on a range of synthetic and real datasets and investigated the applicability of ensemble approaches to gradient boosting models that are themselves ensembles of decision trees. Our analysis shows that ensembles of gradient boosting models successfully detect anomalous inputs while having limited ability to improve the predicted total uncertainty. Importantly, we also propose a concept of a virtual ensemble to get the benefits of...
 

how easy it is to drain the TS on the hours with the spread

 
Maxim Dmitrievsky #:

how easy it is to drain the TS on the hours with the spread

I.e. at spread=7pts it will be 50/50.
And the profitable variant earns only 7 pts per trade on average.
On ECN accounts spread on EURUSD 0-5 usually (average to 3) + ~4 pts for commission. I.e. this strategy will work at 0 on real ECN.
And swaps are now -7.7 and +3.1 pts for some trades will be added for each rollover.
Spread + swap should be taken into account in the markup. Maybe the model will be better, because it will not consider some trades successful during training.

 
Forester #:

I.e. at spread=7pts it will be 50/50.
And the profitable variant earns only 7 pts per trade on average.
On ECN accounts spread on EURUSD 0-5 usually (average to 3) + ~4 pts for commission. I.e. this strategy will work at 0 on real ECN.
And swaps are now -7.7 and +3.1 pts for some trades will be added for each rollover.
Spread + swap should be taken into account in the markup. Maybe the model will be better, because it will not consider some trades successful during training.

and how can the spread be taken into account in the markup if it is deducted from each transaction later, no matter how you mark it up?

 
Maxim Dmitrievsky #:

and how to take into account spread in markup if it is deducted from each transaction later, no matter how you mark it up.

So markup should be based on the financial result. Open-close transaction and transfer the result to the markup. This is the exact variant.

Or subtract the worst variant, for EURUSD on ECN probably 7-10pts, for others maybe more, especially for crosses. + swaps for each day.
On STD accounts it is even worse.

 
Forester #:

So markup should be based on the financial result. Open-close a trade and transfer the result to the markup. This is the exact variant.

Or subtract the worst variant, for EURUSD on ECN probably 7-10pts, for others maybe more, especially for crosses. + swaps for each day.
On STD accounts it is even worse.

I transfer it to markup, after training it still feels bad on the spread

Additionally, I collect a collection of losing trades and teach "do not trade". By the type of bestinterval. Actually, the second meta model does this, as in the articles. It's not very cool either.
 
Maxim Dmitrievsky #: Actually, the second meta model does that, as in the articles. It's not very cool either.

What do you want? We're almost working with randomness. It's not like studying the demand for ice cream depending on temperature, as in the first book on Kozul, which was thrown here six months ago)))))

 
Forester #:

What do you want? We're working almost at random. It's not like researching the demand for ice cream depending on temperature, as in Kozul))).

I want zeekr 001.

 
Maxim Dmitrievsky #:

There's something from Yandex

Thanks, quality and interesting article with extensive literature.

It seems that they do not consider the kind of uncertainty that is interesting - probabilistic dependence of output on attributes. They study two other types of uncertainty - uncertainties related to inaccuracies of attributes and parameters. They are called beautifully - aleatoric and epistemic uncertainty) We should call our variant target uncertainty by analogy).

Imho, in our case "measurement errors" of attributes are absent in principle, and uncertainty of model parameters is poorly separable from our "target uncertainty".

 
Forester #:

What do you want? We're almost working with randomisation. This is not the demand for ice cream depending on temperature, as in the first book on Kozul, which was thrown here six months ago)))))

So we need to try to carefully measure the dependence of this "Almost" on the signs).

1...333733383339334033413342334333443345334633473348334933503351...3548
New comment
Reason: