Machine learning in trading: theory, models, practice and algo-trading - page 3137

 
Maxim Dmitrievsky #:

))))


No, I still think a member rating would be cool.

A couple of times they'd get the fools on the skirting board, and that would be the end of it.
 
Forester #:
What's the point of a trading kozula then? We'll never have reasons for inputs. I'm afraid there won't be associations either.
And the column shuffling was in permutation.
I guess it's supposed to show whether there is a causal relationship between traits and targets or not. I like the proposed approaches, experimenting. Different from the usual learning.
The problem might be the difficulty in isolating the tritment effect due to the influence of confounders in a large feature space. But cross-validation should sort of save the day.
 
mytarmailS #:

Did I offer a grail code?

Practice separates the blabbermouth from the knowledgeable.

You can say whatever you want, there's a lot of them here who have everything simple , but ask one right question and everything tastes good.

Is it clear what we're talking about?


where A,B,C are instruments.

It's almost the same thing:


Switch on your brain first!

You've had a clue drawn to you repeatedly and on all the quid, supposedly one that no one will understand anyway

They're making fun of the human brain.

Doesn't hurt, does it?

Come on, let's see who can solve this charade.

Then we'll draw conclusions.

Response:


 
mytarmailS #:

))))

No, I think a ranking of participants would be cool.

For some reason, an old joke comes to mind:
"Stirlitz stood his ground, it was Mueller's favourite torture" :)
 
mytarmailS #:

it is probably an oscillator of some kind.

Actually what you have done, you decide for yourself why))))

Momentum oscillator

oh cool! thanks! now I understand) only it has to be retrained every time? otherwise you can't recognise components for new data?

 
Evgeni Gavrilovi #:

oh cool! thanks! now I understand) only it has to be retrained every time? otherwise it can't recognise components for new data?

Use umap, not t-sne.

umap has a predictor.


But if the new data goes beyond the range of the old data, the algorithm will not work correctly, in this case it is better to use the usual PCA.

This is all if we are talking about data without normalisation.

 
Renat Akhtyamov #:

so we know what we're talking about?

where A,B,C are instruments

Stop writing heresy here, and also off-topic heresy.

Even rival members of the thread are already echoing in one voice.

 
Maxim Dmitrievsky #:
I guess it is supposed to show whether there is a causal relationship between features and targets or not. I like the proposed approaches, experimenting. Different from the usual learning.
The problem might be the difficulty in isolating the tritment effect due to the influence of confounders in a large feature space. But cross-validation should sort of save the day.
Forester#:
And then what's the point of kozula for trading? We will never have reasons on entry. I'm afraid we'll never have associations either.
And column shuffling was also in permutation.

What does that have to do with your cahuel?

Without being aware of your cajuel, I have been calculating causality for 10 years, estimating it quantitatively, filtering predictors by the variance of the fluctuation of this relationship when the window moves. And I have written a hundred posts about it on this thread.

 
СанСаныч Фоменко #:

What's your cajoling got to do with it?

Not being aware of your cajuel, for 10 years I have been calculating the causal relationship, and evaluating it quantitatively, filtering predictors by the variance of the fluctuation of this relationship when the window moves. And I have written a hundred posts about it on this thread.

how long do you have left? )

Can you take any attributes in sufficient quantity related to time series and any labels showing profits in the tester and make a robust model out of it?

After all, all BP derivatives are relevant to it :)


The task is difficult in other areas where it's not clear at all where the feature comes from and why it's needed. There are tonnes of such rubbish in the big date, which is very difficult to filter. And tonnes of false correlations as a consequence.

Our task looks even primitive compared to this, if we take BP and its derivatives. Because all the signs are related to it.

But we still have to mess with the algorithm and logic to match labels with features. There can be many logics. So you do yours, and we will do ours.

I have already written why I like kozul, because I reached it myself by thinking. And he organically fitted into my idea.

 
For those who say there are no patterns in price: try visiting the field of practising manual traders. Everyone knows sniper or smartmoney (freely available, so not advertising all sorts of schools). The founder after 15 years only managed to convey to his programmers how to algorithmise (and even then not completely) his system. That is, there are working systems and they are related only to the price, and the problem was always the interpretation of squiggles.

And so, a more or less working (raw) Expert Advisor without optimisations fights the sniper's base point in profit all year long. The balance graph is not as beautiful as the traders themselves, but it works. Just a couple of years ago there appeared a methodology with 7 (I think) algorithms of price movement, by which they trade. And if they can "stick" them into the code, they will make an enviable automaton.

What am I getting at? There are regularities in the price. And normal ones, not short-term piss-poor ones. Short-term regularities are when the seller of his grail says that this EA is "tailored" exclusively for EURUSD and next week he will publish a new set, don't miss it! In general, short-term patterns are mostly about fitting. Long-term patterns are algorithms of price movement, targeting big players by their "traces" on the chart, etc. That is, these are patterns that work everywhere, once "adjusted" and it will work on all currency pairs and on all instruments, as long as the eyes are not scattered

If manual traders were able to algorithmise partially trading strategy, then the more neural network should find these patterns, and "increase" them, because neuronka thinks deeper.

That's why I try to cram 500 prices to the input of neural network, so that it "saw" the picture for a month. But, according to this logic, it is necessary to cram not hourly, but minute ones, and there the input increases 60 times, NeuroPro here works with odb databases, and they have a limitation of 512 columns, so I export as it is. And since the programme is ancient, training such a large number of neurons is an eternity.

I tried shoving so many into the input of Python neurons, the output is the same: MLP, LSTM, convolution with two-way LSTM - the result is the same, as if there are no neuron varieties in tensorflow and they are all the same.

In general, don't give up on this case.