Machine learning in trading: theory, models, practice and algo-trading - page 1955

[Maxim Dmitrievsky]

I also wrote a ton of code today, it seems to be interesting

I'll test it tomorrow in the terminal

Still trying to reverse engineer that bot... seems to be already done :D and even better

[Rorschach]

Maxim Dmitrievsky:

subtract from each transaction

You'll lose money, the spread is 10 times bigger

[Rorschach]

Aleksey Vyazmikin:

Is "baseline" a price or a sample? Reset the sample.

Bid increments.

[Rorschach]

in memory

"if the trace and the test are obtained from different distributions (conditionally). That is, the signal that the model learned on the train is simply not present in the test. The simplest way to fight it is to try other validation schemes, by time, by different groups, etc. More complicated - is the so-called adversarial validation, when we first build a model to distinguish the tray from the test, and then use as a validation set that piece of tray, which is most similar to the test."

[Forester]

Rorschach:

in memory

"if the trace and the test are derived from different distributions (conditionally). That is, the signal that the model learned on the train is simply not present in the test. The simplest way to fight it is to try other validation schemes, by time, by different groups, etc. A more complicated one is the so called adversarial validation, when we first build a model to distinguish a tray from a test, and then use as a validation set the piece of tray that is most similar to the test."

As a reminder, I'm saving it to my blog. It's a pain in the ass to find your memo.

I read the article - the idea is interesting. But I think the result will be a fit to the test, not to the trayne. If further new data will be different - again we will get 50/50

[Aleksey Vyazmikin]

elibrarius:
I'm putting it on my blog as a reminder. It's hard to find your memos later.

I read the article - the idea is interesting. But I think the result will be a fitting not for the track, but for the test. If further new data will be different - again we will get 50/50

I wanted to say that not the data will be different, but the outcome, but now I thought, what if you really add information about the surrounding data in samples, which were not involved in the construction of the model? If we just use the tree as an example, then before the last split, we look at statistics for other predictors, identify those that show a statistically positive variant of classification and look to ensure that in the overall sample these predictors are not correlated with the predictor in the last split, then we have additional information about market conditions, which is not taken into account when selecting the last split. Trees, on the other hand, are built according to the greed principle and there is certainly a competition among the predictors on the last split.

Then our split won't just be A>X, but A>X1 && B>X2&& C>X3- i.e. we'll take environment information into account.

[Aleksey Vyazmikin]

What are the analog calculations in the article - who gets it?

[Renat Akhtyamov]

Aleksey Vyazmikin:
What are the analog calculations in the article - who understood?

I can't say for sure, but I suspect that an analog number is immediately a value, and the digital one needs to be converted into this number.

Except that the processor is warm and I'm afraid that their analogs will float over time (give an error) and there will be a lot of interference in addition

[Mihail Marchukajtes]

Renat Akhtyamov:

I can't say for sure, but I suspect that the analog number is immediately a value, and the digital number must be converted to this number.

Except that the processor gets warm and I'm afraid that their analogs will float over time (give an error) and there will be a lot of interference in addition

A little different, digital computers work with information in the form of 0 and 1. Is there current in the conductor or not forming arrays of higher levels. An analog signal can be converted to digital by descritization. But I think what happens here is not addition and mathematical operations with the fact of the presence of current in the conductor or lack thereof, but with the voltage level in a given conductor, which is the analog signal that they have somehow learned to add up. Well stupidly if you want the number five, that's 5 volts on the processor leg, if you want the number 8.345354346346, that's the same number of volts.

As I said in my video, there are two achilles' heels that are not currently realized with respect to the biological neuron. The first is to be able to cut off and make connections between neurons and the second is to work digitally, where there is a desCRIption in the form of a decimal order of magnitude. In the analog world, these dimensions are much larger and I think are individual for each neuron, that is, the analog signal has infinite dimensions, where we can refine the number by an infinite number of digits after the decimal point. The world of real numbers is like that :-)

If I understand correctly and the processor will work with analog signals then this is a real breakthrough where one heel of achilles is already annihilated.... In fact inside our heads neuronchiks transfer among themselves signals in an analog form.

[Andrey Dik]

Aleksey Vyazmikin:
What are the analog calculations in the article - who understood?

Something like tube transistors, it seems, only very small.