Machine learning in trading: theory, models, practice and algo-trading - page 2612

New comment
 
Maxim Dmitrievsky #:

first third of the graph - new data, not involved in training

From the pictures with 25 and 100 iterations you can see that it improved at 100, although the maximum was around 70

Well here I would make a table, 3 options - 1 model, 2 with 25 iterations, 2 with 100 iterations. And some trader metrics (PF, winrate). All on OSS. Somehow "somewhere out there a piece is OSS" and quality metric is apparently one thing for IS+OOS. And purely OOS measured in a human way is something else entirely.

 
Replikant_mih #:

Well here I would make a table, 3 options - 1 model, 2 with 25 iterations, 2 with 100 iterations. And some trader metrics (PF, winrate). All on OSS. Somehow "somewhere out there a piece is OSS" and quality metric is apparently one thing for IS+OOS. And purely OOS measured in a human way is something else entirely.

There are a lot of tables, I just gave you an overview of the working principle.

the idea is not finalised, experiments in the evening with coffee :h
 
Maxim Dmitrievsky #:

lstm always works with MA, tested a long time ago

thanks for the heads up! - forget gate seemed promising, so I decided to start running it... thanks for saving me from a rake) ...

but will still think about how and with what to spin this

 
JeeyCi #:

thanks for the heads up! - forget gate seemed promising, so I thought I'd get started on it... thanks for saving me from a rake)

need some other loss ffs there, otherwise it always feats under MA as the best and easiest option (stupid prediction of past values)

on kaggle somewhere saw exactly the same zafits, they all do it that way )

 
Maxim Dmitrievsky #:

There's a lot of plates, just an overview of how it works.

The idea is not finalised, experiments in the evening with coffee :h

I see, I just don't usually measure anything on the IS when assessing something, that's why it caught my eye.

 
Maxim Dmitrievsky #:

there's a need for some other loss fi, otherwise it always fi ts under MA as the best option and the easiest (stupid prediction of past values)

I've seen the same exact fetches on kaggle somewhere and they all do it )

Is it even if the increment is predicted and not the price?

 
Replikant_mih #:

I see, I just don't usually measure anything on IS when evaluating something, that's why it caught my eye.

Just to see the total number of trades is also interesting, it decreases at each iteration as more and more bad cases are deleted

So far so good.

 
Replikant_mih #:

Is it even if the increment is predicted rather than the price?

I think so )

stationary series for learning, increments, then reconstructed back to the graph
 
Maxim Dmitrievsky #:

there's a need for some other loss fi, otherwise it always fi ts under MA as the best option and the easiest (stupid prediction of past values)

I've seen exactly the same zafits on kaggle somewhere, they all do it that way )

Well, if the predictor is only Close (as tested), there's no other way to teach it... I think it is unlikely that the other way will help (even handwritten by me - I'm not a physicist)) - differentials with vector algebra in handwritten form I can't come up with yet ... available to see
 
secret #:
Here is another example: now gas sales will go for roubles, through intermediaries most likely.
And after that, the seasonal strategy on usdrub related to tax payments by exporters is likely to break down. And it can be meaningfully turned off until the next change in market structure.
And using MO, you have no chance at all of understanding why the incurred system broke and why it worked.

As is easy to understand, it all came down to Timoshenko-inspired gas schemes. And that's the way it is with you (practitioners). You want to talk about your super-exclusive knowledge of the market and the result is either obvious platitudes or something along the lines of the "inevitable dollar collapse" sect or the tales of "experts" on the "wave laws of the market".

1...260526062607260826092610261126122613261426152616261726182619...3659
New comment