Machine learning in trading: theory, models, practice and algo-trading - page 1002

[Грааль]

Alexander_K2:

So, there is an opinion that the sequence of returns (CLOSE[i]-OPEN[i])-(CLOSE[i-1]-OPEN[i-1]) is a stationary series.

One candle's retournee is (close-open)/open, it's clear to hell it's not a clean price to put into NS, the next retournee is predicted by previous ones (with a different window) very poorly, it's not enough for spread, but that's probably all we can get

[Dr. Trader]

Alexander_K2:

In fact, the value CLOSE[i]-OPEN[i] is nothing but the sum of the increments.

The sequence of such values should, in the limit, tend to a normal distribution.

Well, there is an opinion that the sequence of returnees (CLOSE[i]-OPEN[i])-(CLOSE[i-1]-OPEN[i-1]) is a stationary series.

Has anyone tried such a thing on the NS input and what were the results???

Close[i] can be replaced with Open[i+1], in Forex it is true in more than 90% of cases. Or the difference is just a few pips. Then there will be only one time series in the formula, it's more convenient.

Such transformation is used in ARIMA model. And it really serves to achieve stationarity, but there are many more transformations, it's not the only formula there.

https://people.duke.edu/~rnau/411arim.htm

If d=2:  yt  =  (Yt - Yt-1) - (Yt-1 - Yt-2)  =  Yt - 2 Yt-1 + Yt-2

ARIMA is outdated now. In financial markets if it gives something, it is no more than bank interest on a deposit. GARCH is much better, according to the articles, moreover it is ARIMA plus various additions.

[Dr. Trader]

Alexander_K2:

The sequence of such values should, in the limit, tend to a normal distribution.

I haven't seen prices that tend to a normal distribution. My increments always looked like Laplace, with koshi tails.

[Alexander_K2]

This was my theoretical reasoning.

In practice, of course, the first returns do not have Gauss, and no one has ever managed to get it, or will manage to, alas...

But still I was talking about (CLOSE[i]-OPEN[i])-(CLOSE[i-1]-OPEN[i-1]), which is actuallythe second return.

Now, I haven't paid much attention to these second returns so far, and in vain.

[Alexander_K2]

And Kolmogorov, in general, I see, paid special attention to B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i-to])] and refused to predict anything unless this function had a quite definite form.

Maybe it makes sense to put certain conditions on the NS work?

Say, skipping unsteady pieces of BP, exploring, for example, second returns or B(k)?

[Alexander Ivanov]

Hi!

Dear gurus, have you already created a super-bot?

I'd like to try it on the real world.

[Aleksey Panfilov]

Alexander_K2:

And Kolmogorov, in general, I see, paid special attention to B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i-to])] and refused to predict anything unless this function had a quite definite form.

Maybe it makes sense to put certain conditions on the NS work?

Say, skipping unsteady pieces of BP by exploring second returns or B(k), for example?

So there is a limit: (Sigmas squared)

Determining this limit is the first of the problems solved in this

problem solved in this paper.

As for the interpolation problem, we will consider only the

the case of estimating x (/) by the quantities

-x{t + i)Jx{t + 2)1 ...,x(t + n),

x(t - l), x(t~2), ... , x(t - ha).

For this case we denote by oj (ha) the minimum value of the mathematical

of mathematical expectation

а2 = МИ0-<?)%

where Q is a linear form:

Q = axx {t + i) + atx {t + 2)+ ... +apx {t + n) +

+ a-ix(t - l)-\-a-2%(t - 2)+ ... -\-a-nx(t - га)

with constant real coefficients as.

As ha increases, the value of a2 (i) does not increase. Therefore there exists

limit

l im a} (ha) = o? (5)

P~>o

Our second problem is to determine a]. The following proposed

solution of the two problems formulated above has been reported without

proof in my note (*) *. It relies on concepts related to the

to the spectral theory of stationary random processes.

The spectral theory of stationary random processes was

constructed by A. Я. Hinchin for the case of continuous changes of

argument t (2 ) .

I do not understand, do you plan to analytically estimate the reliability of the prediction already made, or to make a prediction to begin with. The first couple of pages say that the article is about assessing the reliability of the prediction. And the forecasts themselves should be found inA.J. Hinchin.

And you haven't carefully copied the basic statement from the article.

Not: B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i-to])

A: B(k)=M[x(t)*x(t-to)]=M[(CLOSE[i]-OPEN[i])*(CLOSE[i-to]-OPEN[i])

In addition, I think it is more correct:

Forum on trading, automated trading systems and testing trading strategies

Machine Learning in Trading: Theory and Practice (Trading and Beyond)

Dr. Trader, 2018.07.06 02:37

Close[i] can be replaced by Open[i+1], in forex it is true in more than 90% of cases. Or the difference in a couple of pips only. Then there will be only one time series in formula, it is more convenient.

Such transformation is used in ARIMA model. And it does serve to achieve stationarity, but there are many more transformations there, it's not the only formula there.

https://people.duke.edu/~rnau/411arim.htm

If d=2:  yt  =  (Yt - Yt-1) - (Yt-1 - Yt-2)  =  Yt - 2 Yt-1 + Yt-2

ARIMA is already outdated, on financial markets if it gives anything, it is no more than bank interest on a deposit. GARCH is much better, according to the articles.

PS.

Oh yes, and thanks for the reply to my question from my post: https://dxdy.ru/post1244134.html#p1244134

[Gramazeka1]

Hello, this is Misha speaking and as you guessed I'm doing it from the phone :-)

In general, I think that pre-processing of data comes to the forefront now. Of course, the optimization algorithm is also important in itself, but the availability of a good training sample is also not the last thing. You all slandered and slandered Reshetov's optimizer but by the way it makes good models if data are well pre-processed. In any case with 10 optimizations at least half of the models will work. After all, his implementation is not so simple. And I think JPrediction will always be relevant. The main thing here is to correctly preprocess data and it is in this area now the competition.

[Gramazeka1]

As I said, in another thread .... Now I'm working on a package for p and at least two preprocessing already implemented. The first removes garbage predictors. The second makes the sample more presentable for training and here on the second point I would like to stop.

After the representativeness was done, the quality of learning increased by 15 percent for the same data set. The main task is to increase the learning period while maintaining the proper level of learning quality. Example: With 25 examples, I could get from 80 percent generalizability to yours. By processing the representativeness on 40 examples I was able to get 90% of the qualities of the model. Based on the postulate that the best model is the one that was able to learn over a longer sample while preserving the quality of learning.

[Alexander_K2]

Gramazeka1:
Hello, this is Misha speaking to you and as you guessed I do it from the phone :-)

In general, I think that pre-processing of data comes to the forefront now. Of course, the optimization algorithm is also important in itself, but the availability of a good training sample is also not the last thing. You all slandered and slandered Reshetov's optimizer but by the way it makes good models if data are well pre-processed. In any case with 10 optimizations at least half of the models will work. After all, his implementation is not so simple. And I think JPrediction will always be relevant. The main thing here is to correctly preprocess data and that's where the competition is now.

Hi Misha!

Yes, it's time to reconsider all neural networking efforts and their meager hopes for the tool itself. Nothing will help - neither forests, nor steppes - if the input data are not prepared.

And yes - there is no competition, there is a problem and there is general dumbing down.

If you know how to prepare the data, lay it out. Humanity will thank you.