Machine learning in trading: theory, models, practice and algo-trading - page 3429

[fxsaber]

Please give me your thoughts on the topic of self-deception.

1. I did Optimisation by Genetics (interrupted) on Sample, sorted the results by optimisation criterion, leaving only the top 50.
2. Out of these 50, I chose 2-3 that pass OOS left and right perfectly. Length_OOS = 2 * Length_Sample.
   

How strong is the fit factor in the second point?

[mytarmailS]

fxsaber optimisation criterion, leaving only the top 50.

Out of these 50, I chose 2-3 that pass OOS left and right perfectly. Length_OOS = 2 * Length_Sample.

How strong is the fit factor in the second point?

100%.

[Andrey Dik]

fxsaber optimisation criterion, leaving only the top 50.

Out of these 50, I chose 2-3 that pass OOS left and right perfectly. Length_OOS = 2 * Length_Sample.

How strong is the fit factor in the second point?

The greater the differences in the parameters of these 3 sets, the less likely the fit is.

[fxsaber]

Andrey Dik #:

The greater the differences in the parameters of these 3 sets, the less likely the fit is.

Very interesting statement, thank you. True, it was understood a bit differently.

The greater the differences in the parameters of the 50 sets (genetics is specially interrupted for this purpose), the lower is the probability of fitting in point 2.

[mytarmailS]

))

[Maxim Dmitrievsky]

mytarmailS #:

you don't get it.

Okay, do it.

The toaster needs to be rewritten for multicurrency first.

Then I'll add an option to embed fractals.

Then I will add the option of kozul via cross-learning. So fractals are only a detail.

That is, you throw in different pairs for better generalisation, and add more patterns to the overall pot.

If treebanks don't work well with such "embeddings", I might consider a neural network.

Then I'll switch to LLM. Mbe will need a new comp with a new graphics card. Or I'll have to buy a subscription to colab to get more power.

[mytarmailS]

Maxim Dmitrievsky #:

The toaster has to be rewritten for multicurrency first.

Then I will add an option to embed fractals.

Then I will add the option of kozul via cross-learning. So fractals are only a detail.

That is, you throw in different pairs for better generalisation, and add more patterns to the overall pot.

If treebanks don't work well with such "embeddings", I might consider a neural network.

Then I'll switch to LLM. Mbe will need a new comp with a new graphics card. Or I'll have to somehow buy a subscription to colab to get more power.

That's not ebding.)

so colab gives you a free graphics card, it should be enough for your needs.

[Maxim Dmitrievsky]

mytarmailS #:

it's not yembedding.)

so colab gives you a vid for free, it should be enough for your needs.

It may not be enough for "LLM", they are very resource-hungry.

Embedding is conditional, I don't know what to call it otherwise.

[mytarmailS]

I need a simple but very fast backtester written in C++ or rust with support for limit entry, stops and takers.

if anyone knows such a backtester, please advise.

[fxsaber]

mytarmailS #:

I need a simple but very fast backtester written in C++ or rust with support for limit entry, stops and takes.

if anyone knows such a backtester, please advise.

Virtual is written in MQL5. It should be ported to C++ without any problems. I think it is one of the fastest tick backtesters in the world.