Machine learning in trading: theory, models, practice and algo-trading - page 2747

[СанСаныч Фоменко]

It was slipped in above that the value of even a correct prediction can be zero.

To me, the value of a prediction depends not only on the probability of error, but on the teacher himself.

Is the teacher predicting the next bar or the trend?

What time period is the prediction being made and what time period are we trading on?

What target profit is our teacher predicting?

What is the relationship between the teacher and the spread?

There are a lot of questions that have not been discussed here.

[Valeriy Yastremskiy]

СанСаныч Фоменко #:

It's principled, principled rewiring. We don't need models that live for 100 years. We need a model that will give a prediction for the next bar with low error. Then comes the advisor, and there are their own problems with this prediction.

I don't get it at all. What is the logic of the EA?

[СанСаныч Фоменко]

Valeriy Yastremskiy #:
I don't get it at all. What is the logic of the counsellor then?

The logic of the advisor is determined by the teacher.

But re-learning on each new bar is fundamental. We are moving away from the concept of model "life" time, as there is no "life" on non-stationary markets. Because of non-stationarity, all these tests "out-of-sample" are meaningless.

[mytarmailS]

So what's the error on the new data?

Otherwise, what's the point of all this

[СанСаныч Фоменко]

mytarmailS #:
So what's the error on the new data?

Otherwise, what's the point of all this

Once again: no more than 20%. There was more detail above.

[mytarmailS]

СанСаныч Фоменко #:

Once again: no more than 20 per cent. It was more detailed above.

I remember, but it's a fitting error, I mean, trace sampling....

What's the error of that next candle, the test.

[Maxim Dmitrievsky]

СанСаныч Фоменко #:

It's principled, principled rewiring. We don't need models that live for 100 years. We need a model that will give a prediction for the next bar with low error. Then comes the Expert Advisor, and it has its own problems with this prediction.

Well, it makes a certain sense, because often models do not live long. But I would like to find options without retraining constant, at least on the interval of a year +, with a slow degradation of the model, which is easy to trace

I've written one-armed bandits and bots on reinforcement learning. In essence, they suffer from the same disease - either they adapt slowly, or too frequent rebuilds lead to accumulation of error.

I always put the spread in the models. Actually it kills them, without spread and commission would work.

Even if clustering of increments with 1-5 lags is done, without spread it works +. That is, the main fish is at the micro level. But it's understandable. The fact that they are always banned as toxic TCs.

[Valeriy Yastremskiy]

СанСаныч Фоменко #:

The logic of the counsellor is determined by the teacher.

But repeated training on each new bar is fundamental. We are getting away from the concept of model "life" time, as there is no "life" in non-stationary markets. Because of non-stationarity, all these tests "out-of-sample" are meaningless.

I don't argue with training on every bar, and maybe even on a non-stationary tick for example. I don't understand the structure of training then completely. Is EA logic a separate training or part of the training on each bar? It is like the tails of the first training, how many tails, or stages of training?

[Valeriy Yastremskiy]

On topic)

Forum on trading, automated trading systems and testing trading strategies

Not the Grail, just ordinary such - Bablokos!!!!

Aleksey Nikolayev, 2022.09.16 16:28

Well, mathematicians are very strong in inventing new signs and new values for old ones. The main problem in reading mathematical texts is to assimilate the system of notations, which is different for each author. And the texts of theoretical physicists are impossible to read at all - they are complete outlaws in the sense of originality of used notations. Therefore, if there is any "true" meaning of some mathematical notation, it is not known to anyone).

[Renat Fatkhullin]

Please note that from build 3440 we start distributing AVX versions of the software: https://www.mql5.com/ru/forum/432624/page5#comment_42117241.

The next step is to rewrite the mathematical apparatus to vector and OpenCL functions, which gives ten-tract accelerations without the need to install additional libraries like CUDA.