Bayesian regression - Has anyone made an EA using this algorithm? - page 5

 
It seems a bit immodest that the only one who makes any sense here is gpwr. The rest of you, just excuse me:)
 

)))

And I get the impression that someone simply has nothing better to do and start getting clever, although the opinion is - no go) ... and as you can see above - not only mine

 
Alexey Burnakov:

One should use a method in which the density of the error distribution is not important. Non-parametric methods.

We do not know the error distribution for forex at all. Formally - and strictly - errors are differences between modelled values and modelled values obtained on the gene population, i.e. purely theoretical values. Residuals are obtained on distinctions of modeled values from model values on the available sample, but they will hardly be normal as well, as financial time series (their returns, to be more exact) are not normal (!) and are thickly lined and peaked, while it is very difficult to model such athickly lined and peaked series.

I even bothered and derived for hourly increments the original distribution (turquoise =)) and the normal one with the same mean & sd parameters. As you can see it is far from being normal. And the normality test is far from passing.

Methods that rely on normality of errors are classical, from the 20th century, methods such as linear regression and analysis of variance. But we can do without them.

Read the wiki.)

If you conducted researches at flat, as the authors of bitcoin strategy did, then you surely know better how the differences between the real and ideal curve affect the result.

The Gaussian distribution, the most popular in nature and widely used in science (from sociology to nuclear physics) is accepted with hostility by many in the MQL community regarding its applicability to the market.

I'm not a mathematician, but when I look at the distribution of bars or tick volumes by price levels, the picture reminds me of a bell. Especially on the flat markets. For example. The whole EURUSD story looks like a global flat.

 
Yuri Evseenkov:

If you conducted researches at flat, as the bitcoin strategy authors do, then you surely know better how differences between the real and ideal curve affect the result.

Gauss distribution is the most widespread in nature and widely used in science (from sociology to nuclear physics) for some reason it is not accepted by many in MQL community as applicable to the market.

I'm not a mathematician, but when I look at the distribution of bars or tick volumes by price levels, the picture reminds me of a bell. Especially on the flat markets. For example. The whole EURUSD story looks like a global flat.

Density is measured on price increments, not prices themselves.
 
Alexey Burnakov:
Density is measured on price increments, not prices themselves.
О! Now that's interesting. Can I have the formula?
 
new-rena:
Oh! Now that's interesting. Can I have the formula?

Colleague, these are the basics!

You can take different formulas, such as the most popular ones:

Pr - price

t - time

1) Pr(t) - Pr(t-1)

2) Pr(t) / Pr(t - 1) - 1

3) log(Pr(t)) - log(Pr(t-1))

So when economists say that we have measured, for example, the variance of such an instrument, they do the following: variance = sum((Xi - X^)^2) / (N - 1),

where Xi is the increment calculated by one of the formulas,

X^ is the X with a cap - the sample estimate of the mean incremental value in the available sample

N - 1 is the sample size minus one,

and the whole formula is an unbiased estimate of the variance.

And then these economists start thinking that the density of increments is normal and try to do a thing like: sqrt(variance) * sqrt(m) * 1.96,

where the root of variance is an estimate of standard deviation and the entire formula is a stretching of the consequence of normality on the non(!)normal series to get an estimate of the extreme limit of price spread in m steps forward with 95% probability. And errors are obtained, of course.

I hope I've explained approximately. And the initial price series doesn't resemble a normal one even in the first approximation, unlike the increments.

 
Alexey Burnakov:

Colleague, these are the basics!

You can take different formulas, such as the most popular ones:

Pr - price

t - time

1) Pr(t) - Pr(t-1)

2) Pr(t) / Pr(t - 1) - 1

3) log(Pr(t)) - log(Pr(t-1))

So when economists say that we have measured, for example, the variance of such an instrument, they do the following: variance = (Xi - X^)^2 / (N - 1),

where Xi is the increment calculated by one of the formulas,

X^ is the X with a cap - a sample estimate of the mean value of the increments in the available sample

N - 1 is the sample size minus one,

and the whole formula is an unbiased estimate of the variance.

And then these economists start thinking that the density of increments is normal and try to do a thing like: sqrt(variance) * sqrt(m) * 1.96,

where the root of variance is an estimate of standard deviation and the whole formula is a stretching of the consequence of normality on the non(!)normal series in order to get an estimate of the extreme limit of price spread in m steps forward with 95% probability. And errors are obtained, of course.

I hope I've explained approximately. And the initial price series does not resemble a normal one even in the first approximation, unlike the increments.

I looked at the formulas. Yes, this approach fits here. Thank you!

I want to read the basics. Maybe there's a textbook on the above subject?

 
new-rena:

I looked at the formulas. Yes, it's glued to this approach. Thank you!

I want to read the basics. Maybe there's a textbook on the above topics?

There's a good earful of the basics

Лекция 14: Линейная регрессия и корреляция
Лекция 14: Линейная регрессия и корреляция
  • 2014.01.29
  • www.youtube.com
Излагается метод линейной регрессии. Лекция и тесты в НОУ ИНТУИТ http://www.intuit.ru/studies/courses/637/493/lecture/11167
 
new-rena:

I looked at the formulas. Yes, it's glued to this approach. Thank you!

I want to read the basics. Maybe there's a textbook with the above subject matter?

Honestly, I haven't read any textbooks myself. Basically, I catch on in the process of analysis.

The main thing in this case, do not take for granted the words of scholars. I tell you, still stock analysts take them as a normal process simply because it is convenient.

I would recommend a book on time series analysis. But there will also be a bunch of Arima, Garch, Unit Root stuff in there that might not apply to forex at all.

 
Alexey Burnakov:

variance = sum((Xi - X^)^2) / (N - 1),


According to this formula on the trend the variance will be 0. Is this the right one?