Bayesian regression - Has anyone made an EA using this algorithm? - page 42

 
Alexey Burnakov:

Brief distribution analysis in R:

We estimated the parameters of the normal distribution from the available clock-bar opening price increments and plotted to compare the frequency and density for the original series and the normal series with the same distributions. As you can see even by eye, the original series of increments of hour bars is far from being normal.

And by the way, we are not in a temple of God. It is not necessary and even harmful to believe.

Write the function F(x) = a*exp(-b*|x|^p) into your distribution. p=2 will give a normal distribution. When you know the true value of p, replace the minimization of the sum of squares of the regression errors with the sum |error|^p. I've shown the output before in this thread. If you think that minimizing the sum |error|^p will give you better prediction accuracy than minimizing the error^2 sum, then go ahead and implement it.
 
Alexey Burnakov:

First I would like to see a glimmer of understanding in the eyes of the 'faithful'. And then, yes, convert if necessary. Whether thick tails can be converted is the question. They can make a big difference to quality.

I'm afraid to repeat myself, but converting thick tails is not an issue.

What kind of quality do you think it would affect?

https://www.mql5.com/ru/forum/72329/page14#comment_2253485

Bayesian regression - Делал ли кто советник по этому алгоритму?
Bayesian regression - Делал ли кто советник по этому алгоритму?
  • www.mql5.com
Bayesian regression - Делал ли кто советник по этому алгоритму? - Страница 14 - Категория: автоматические торговые системы
 
Alexey Burnakov:

It's the same thing! The increments turned into + or - signs. And you can take such a sign for increments one hour ahead.

What is the question?

I have a classification model: learning Buy, Sell. Evaluating the model by coincidence/non-coincidence correctness of direction

An increment e.g. greater than zero is not necessarily a Buy as the increment has a confidence interval. And the evaluation is an error, e.g. MAE

 
Vladimir:
Write the function F(x) = a*exp(-b*|x|^p) into your distribution. p=2 will give a normal distribution.
The thinking is simply revolutionary. Here on the site it's laid out in a nutshell. Approximating to normal is quite possible, but get ....
 
СанСаныч Фоменко:
The thought is simply revolutionary. It's shelved here on the website. Getting close to normal is quite possible, but get ....
You can also do it with Box Cox if the distribution of series deviations is known in advance and static. I think people here are confused by two important things: the distribution of regression errors and the distribution of the input series itself. RMS regression does not care how the input is distributed. The main assumption is that the distribution of model fit errors must be normal. Again, if you don't like RMS regression with its normal ERROR requirement, then use general regression with "non-normal" errors |error|^p.
 
Vladimir:
You can also use Box Cox if the distribution of series deviations is known in advance and static. I think people are confusing two important things here: the distribution of regression errors and the distribution of the input series itself. RMS regression does not care how the input is distributed. The main assumption is that the distribution of model fit errors must be normal. Again, if you don't like RMS regression with its normal ERROR requirement, then use general regression with "non-normal" errors |error|^p.

For some reason, I am fully convinced that the requirement of stationarity of input variables is crucial for deciding the applicability of regression analysis in principle. The whole idea of ARMA is built on a discussion of the stationarity of precisely the input variables, with their non-stationarity transformed to a stationary form by differentiation in ARIMA models. In all this, there are serious difficulties in proving the stationarity property of the time series itself.

As for the regression fitting error, this is from the realm of stationarity. While differentiating the time series makes it possible to practically remove variability in the mean, variability in the variance is dealt with by the ARCH tool.

It is so detailed, as it is absolutely unclear how thousands and thousands of very competent people could not find such a simple means to combat non-stationarity of a time series and it turns out that there is a RMS regression, which solves all the problems with stationarity, which have been studied since about the mid-70s.

 
СанСаныч Фоменко:

For some reason, I am fully convinced that the stationarity requirement for the input variables is crucial to deciding whether regression analysis is applicable in principle.

Non-stationary data are not predicted by time series models. Neither statistical models (regression, autoregression, smoothing, etc.) nor structural models (NS, classification, Markov chains, etc.).

Only subject area models

 
Vladimir:
Write the function F(x) = a*exp(-b*|x|^p) into your distribution. p=2 will give a normal distribution. When you know the true value of p, replace the minimization of the sum of squares of the regression errors with the sum |error|^p. I've shown the output before in this thread. If you think that minimizing the sum |error|^p will give you better prediction accuracy than minimizing the error^2 sum, then go ahead and implement it.
Maybe I'll try a conversion. Thanks
 
СанСаныч Фоменко:

For some reason, I am fully convinced that the requirement of stationarity of input variables is crucial for deciding the applicability of regression analysis in principle. The whole idea of ARMA is built on a discussion of the stationarity of precisely the input variables, with their non-stationarity transformed to a stationary form by differentiation in ARIMA models. In all this, there are serious difficulties in proving the stationarity property of the time series itself.

As for the regression fitting error, this is from the realm of stationarity. While differentiating the time series makes it possible to practically remove variability in the mean, variability in the variance is dealt with by the ARCH tool.

It is so detailed, because it is not clear how thousands and thousands of very competent people could not find such a simple means to combat non-stationarity of the time series and it turns out that there is a RMS regression, which solves all the problems with stationarity, which are studied since about the middle of the 70s.

Please explain finally (someone, or better all at once) what you call stationarity, how do you understand it?
 
Dmitry Fedoseev:
Please explain finally (someone, or better all at once) what you call stationarity, how do you understand it?
Stationarity is the property of a process not to change its characteristics over time.