Regression equation - page 8
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
What's wrong with a zigzag to find the minimum of a function?
and how are you going to build a zigzag, for example, in ten-dimensional space? )))))) And how many computational resources will it take?
and how are you going to build a zigzag, for example, in ten-dimensional space? )))))) And how many computational resources will it take?
I was interested in the regression equations. However, I have encountered a problem describing them adequately. What data do we have: time (say M15), HIGH, LOW, OPEN, CLOSE, VOLUME. For us it is a set of observations. We have an indicator for which we need to establish a functional relationship with the object parameters (in our case, the change in the exchange rate) - factors. Required: to establish a quantitative relationship between the indicator and the factors. In this case, the task of regression analysis is understood as the task of identifying the functional dependence y* = f(x 1, x 2, ..., x t) that best describes the data we have.
The function f(x 1, x 2, ..., x t) that describes the dependence of the indicator on the parameters is called regression equation (function).
So. Question 1: Of the data we have, which one should we choose as the Indicator and which one should we choose as the Factor? Logically the Indicator is time, the factors are H, L, O, C, V
In our case it is a time series.
The next task is to choose the functional dependency. An equation that describes the relationship between the variation of the indicator and the variation of the factors. Often these are polynomial functions. A particular case is the 1st degree polynomial - a linear regression equation.
Question 2: What is the best polynomial to choose, and how to adequately describe it in terms of time series, what parameters to apply, what is the degree of the polynomial. Has anyone used Chebyshev polynomial? If so, what is the order?
Our next task is to calculate the coefficients of the regression equation. The usual way is to use ANC.
Question 3: What is the best method for calculating the coefficients for our case?
Question 4. Do you need to normalize the data?
The topic is certainly an important and interesting one.
So we have a time series containing N samples. At this stage it's not important what exactly should be understood as samples - ticks, OHLC or something else. What seems important is the answer to the question about the optimal training sample length n not equal to N, the optimal number of adjustable parameters k<=n (degree of polynomial) and the prediction horizon T (measured in counts).
At this stage the particular type of approximating function and the method of its approximation to the original series are not important. It is important to obtain the dependences of the above parameters on the properties of the initial BP. It is known, for example, that if BP is an integrated random variable, then the optimal prediction is a constant equal to the value of the last reading (zero bar). If the series contains regularities, one must look for the optimum in terms of regression parameters.
Any common sense considerations in this setting?
The topic is certainly an important and interesting one.
So we have a time series containing N counts. At this stage it doesn't matter what exactly is meant by counts - ticks, OHLC or something else. What seems important is the answer to the question about optimal training sample length n not equal to N, optimal number of adjustable parameters k<=n (polynomial degree) and forecast horizon T (measured in counts).
At this stage it is not important the particular type of approximating function and the method of its approximation to the original series. It is important to obtain the dependences of the above parameters on the properties of the initial BP. It is known, for example, that if BP is an integrated random variable, then the optimal prediction is a constant equal to the value of the last reading (zero bar). If the series contains regularities, one must look for an optimum by regression parameters.
Any common sense considerations in this formulation?
No shit in this formulation. Fucking theorizing. Kick this polynomial regression to the wall with one BP.
We need profit maximisation. Fuck all univariate regressions. Why use only a fraction of the market information? When there is plenty of information.
Regression analysis should be multivariate, that's the minimum. Analysis of different estimation methods (IOC, MO of absolute error values (Laplace or Lagrange - can't remember), sign, quantile, etc.) of regression on efficiency.
Estimating forecast horizon is also an interesting song.
Wrote some crap on the subject. There's not much there, of course. Only the very beginning. Ahead of that is estimating the BP profit forecast horizon and lots of interesting bummer...
Wrote some shit on the subject.
What are you so excited about?
Do you think the more you pile up the different, preferably not simple and transparent, will be better?
Experience tells a different story. It is right - simpler, and to understand thoroughly in the studied subject properly! And "multivariate regression", "quantile"... - It's like spinor analysis of torsional interaction.
Shit, I didn't throw anything, where did you get that from. I have a simple multivariate LINEAR regression in general. And the rationale for using linear regression lies in the logic of making an optimal portfolio and finding correlations. That is the starting point - from simple.
Fuck knows what you mean by regression, found out what it is myself the other day. I mean regression analysis.
... I have a simple multivariate LINEAR regression...
but you can do a multivariate polynomial regression... is it worse than linear ? i don't know, there is only one check - if prediction accuracy increases or prediction time increases with the same accuracy, then yes it's better ... But to check it you don't just have to understand how to do it, you also have to explain it to the machine ...
but you can do a multivariate polynomial regression ... I don't know, there is only one check - if prediction accuracy increases, or prediction time increases with same accuracy, then yes it's better... But to check it you don't just have to understand how to do it, you also have to explain it to the machine ...