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

 

The strategy is able to nearly double the investment in less than 60 day period when running against real data trace.

I. Bayesian Regression The problem. We consider the question of regression: we are given n training labeled data points (xi,yi) for 1 ≤ i ≤ n with xi ∈ Rd,yi ∈ R for some fixed d ≥ 1. The goal is to use this training data to predict the unknown label y ∈ R for given x ∈ Rd. The classical approach. A standard approach from non-parametric statistics (cf. see [3] for example) is to assume a model of the following type: the labeled data is generated in accordance with relation y = f(x)+ where is an independent random variable representing noise, usually assumed to be Gaussian with mean 0 and (normalized) variance 1. The regression methodboilsdowntoestimating f from n observation (x1,y1),...,(xn,yn)andusingitforfutureprediction. For example, if f(x) = xTθ∗, i.e. f is assumed to be a linear function, then the classical least-squares estimate is used for estimating θ∗ or f: ˆ θLS ∈argmin θ∈Rd n X i=1 (yi -xT i θ)2 (1) [...]Bayesian regression and Bitcoin.pdf
 

First of all a normal exchange was used, second probably the main data that was fed was the market cup, plus you can get separate volumes there, and thirdly it was cryptocurrency.

The Pdf is bearded, I don't think it's that easy now, it's much harder to trade.

But I think it is still possible to pinch.

If you want to do such a thing with forex, drink cold water, I'll tell you right away - it will not work with 99% probability.

 
Комбинатор:

First of all a normal exchange was used, second, probably the main data that was fed is a market rate, plus there you can get separate volumes, and third, it is a cryptocurrency.

The Pdf is bearded, I don't think it's that easy now, it's much harder to trade.

But I think it's still possible to pinch.

If you want to do such a thing with forex, drink cold water, I can tell you right now - it will not work with 99% probability.

Thank you.

Interested in the opinion on whether this strategy can be used in forex. I have a lot of different opinions, especially precious ones, which are based on practical experience.

 
lilita bogachkova:

Opinions may vary, especially those based on practical experience are valuable.
In fact, it's simple - if you can find data on which the regression will find a pattern, you'll get it. But the ticks in Forex are aggregated smoothed price changes without any additional information. Do you think it will work with this kind of data?
 
lilita bogachkova:

Thank you.

I am interested in the opinion on the possibility of using this strategy in forex. Opinions may vary, especially those based on practical experience are valuable.

Bayesian regression is similar to ridge regression, but is based on the assumption that the noise (error) in the data is normally distributed - therefore it is assumed that a general understanding of the data structure is already available, and this makes it possible to obtain a more accurate model (compared to linear regression for sure).

Hence: http://datareview.info/article/10-tipov-regressii-kakoy-vyibrat/

Due to the assumption of normality of errors, I question the applicability of this method to financial markets.

In addition, in any model where the dependence is only estimated as a hyperplane, there is a chance of missing a non-linear edge, which is what can make the model profitable.

10 типов регрессии – какой выбрать?
10 типов регрессии – какой выбрать?
  • votes: 4
  • datareview.info
Сегодня мы расскажем о десяти основных видах регрессии и подскажем, какой из них выбрать исходя из контекста поставленной задачи.
 
Alexey Burnakov:

Due to the assumption of normality of errors, I question the applicability of this method to financial markets.

Why? Do you think that the law of normal distribution does not work in financial markets? If you consider noise (errors) to be random quantities, then it is appropriate to use Gaussian distribution for them.
 
It really doesn't depend much on the regression, more on the input data.
 
Комбинатор:
It really doesn't depend much on the regression, more on the input data.

I would even reinforce your point: the success of modelling is not determined by the models (if applied correctly), but by the input data. In my experience, some input data sets ALWAYS lead to over-trained models, and if you work with them, you may well end up with un-trained models. I believe that the main evil of modelling - overfitting - is determined by the input data. Solving this problem with regularisation is a half-measure.

To the most promising ones I refer: ada, randomforest, SVM.

Bayesian model because of the assumption of normality of the error (as well as many other models) is better not to use.

 
СанСаныч Фоменко:

I would even reinforce your point: the success of modelling is not determined by the models (if applied correctly), but by the input data. In my experience, some input data sets ALWAYS lead to over-trained models, and if you work with them, you may well find that you can get models that are not over-trained. I believe that the main evil of modelling - overfitting - is determined by the input data. Solving this problem with regularisation is a half-measure.

To the most promising ones I refer: ada, randomforest, SVM.

Bayesian model because of the assumption of normality of the error (as well as many other models) is better not to use.

Thank you.

I am also interested in the opinion on the trading approach described in the article.

 
On the wrong side of things. You have to take it, code it and check it out. You're already starting to... ...speculating about distribution normality.
 
Yuri Evseenkov:
Why? Do you think that the law of normal distribution does not work in financial markets? If you consider noise( errors) to be random quantities, then it is perfectly appropriate to use Gaussian distribution for them.

A situation where the errors will actually be normal is rare, and this requires a careful reproduction of the probability density function for the original series in the model. Will this be achievable? That is the question. And if the plausibility of the model parameter estimates depends on it, you can miss. I would use non-parametric methods, the same random forest, GBM, non-linear SVM.

But generally speaking people that understand linear regression well and can do feature engineering have gotten better results in financial markets than random guessing.