Machine learning in trading: theory, models, practice and algo-trading - page 2757
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
Correct. For lack of a priori assumptions, the second type is used.
In my view, there are two types of connection.
The first one is causal, which is determined by a priori information about the object of research from knowledge in the given subject area, rather than by some calculations.
The second type is probabilistic dependence, which can be calculated a posteriori from some data obtained by observing the behaviour of the object. The second type includes correlation, deterministic dependence (as an extreme case) and so on and so forth, including that described by copulas and other methods. The basis for studying this type is the assumption that there is a joint distribution for predictors and target.
For lack of experiments the 2nd type is used (e.g. US Food & Drugs Association - will not test a normal representative sample for its conclusions, so it relies on Bayesian approaches)... and without a priori information, there is nothing to model at all
For lack of experiments , the 2nd type is used (e.g. US Food & Drugs Association - will not test a normal representative sample for its conclusions, so it relies on Bayesian approaches)... and without a priori information, there is nothing to model at all
Has anyone participated in the Numerai competition? What do you have to do to earn money there?
Do you have to invest your own money? I don't understand what their payout model is.
Has anyone participated in the Numerai competition? What do you have to do to earn money there?
Do you have to invest your own money? I don't understand what their payout model is.
I haven't looked at the library, the article is disgusting - contradicts the common sense of the stat ....
in standard English sources - the meaning of time series analysis is reduced to the change of policy at the moment of treatment/intervention and the analysis of the change of slope of the aggregate trend (which, I suppose, can be interpreted as an actor -- experiencing the influence of policy and modifying its decision-making-process at the moment of treatment -- which is what marketers' research is aimed at when they evaluate the effect of discounts, sales, etc. promotions in order to figure out whether the price does not suit customers, or the product in principle, or the location of the shopping centre, etc.)....д.)...
but the same problem as always in modelling - to evaluate post-treatment, of course, you need a sample(!) to approximate the conclusions "helped-not helped-indifferent" (in terms of intervention)...
and in terms of counterfactual - it is important to ask the right question to assess the dynamics of changes caused by a policy change (or some intervention) - to choose the metric, target and parameters (for tuning) -- because different questioning can give different results (and different slope change) - hence different conclusions.
I am confused by the imbalance problem in ML real data (which biases estimates) - does anyone here solve it with oversampling/undersampling? -- I don't see the point of distorting the real data in such a way....
but it is necessary to obtain a representative sample at the pre-treatment stage (a priori probability distribution), and the posterior distribution is obtained in post-treatment (e.g., policy change)... this is where it is important to decide on your Stopping Rule - i.e. whether to increase the sample to refine the results or to make do with the chosen sample limit to draw a conclusion, which will probably be less statistically significant than if we increase the sample.... but it is not certain that increasing the sample will increase the statistical significance of the mean or variance.
= this is a size problem ... usually, if the effect of intervention is large, it can be seen in a small sample....
the problem of factors (FS) also remains - by increasing the number of factors considered, we reduce the bias of estimates, but increase the variance ... task: to find significant factors (as usual in Explorative Data Analysis - that's why it's calledData_Science, and not a stupid programmer's approximation of random) to obtain unbiased estimates with low variance (the balance of these two goals is at the developer's discretion).
Vladimir has already stated a lot about the problem of selecting factors - if we are modelling probabilities for selecting a high-probability trade.
P.S..
speed and acceleration (if any) are always important in timeseries analysis, their comparison on pre-treatment and post-treatment period gives conclusions (about change of direction including)...
divergence/convergence and extrema of correctly selected targets also remain valid... everything is as usual - it's all about the Design/Architecture of the neural network... and only trends and probabilities of their development are predicted - nothing more... and in the market for day traders everything changes faster than in a long-term trend (if analysed by D1) - so the time factor should also be put into the robot's model for day trading.... in general, formalise your trading style, so that you don't have to sit in front of the monitor all the time. And, if you like, look for statistical reasons for entries and exits or staying out of the market (even for the reason of risk-management -- when the market is not clear).
p.p.s
the topic can be developed endlessly in terms of studying Structural Causal Models (what depends on what, as I noted earlier) - including consideration of exogenous (influence from outside) and endogenous (e.g. commodity or financial currency, and even change of ruling party, I guess) factors.... in general, as usual, you can examine the data for any hypothesis and look at the acceptance or rejection of the null hypothesis for a particular significance level of interest (increasing the sample size for its [significance level] possible improvement).
p.p.p.s
although some people do not like the word probabilistic distribution - but the essence of it does not change - distributions are still probabilistic, even if they are conditional (the condition gives a reason for classification) ... and Before_treatment and After-treatment (in A/B test) can be considered as a change of conditions (policy), but it is possible to estimate regression or compare variance (whether it has changed), even if the slope is the same.
that's why I'm saying that you should first decide on the algorithm (including imbalances - I don't know what you wanted to do with them ???)... and then look for a lib that allows you to charge the code with the necessary entities/classes... - when you advised oversampling earlier)... and then look for a lib that allows you to add the necessary entities/classes to the code... or code your own library with the necessary classes... or code your own library with the classes you need.
that's why I'm saying that you should first decide on the algorithm (including imbalances - I don't know what you wanted to do with them ???)... and then look for a lib that allows you to charge the code with the necessary entities/classes... - when you advised oversampling earlier)... and then look for a lib that allows you to add the necessary entities/classes to the code... or code your own library with the necessary classes... or code your own library with the classes you need.
Everything you need has been coded before you.
The caret shell from R contains up to 200(!) models, in your terminology (libraries) + all the necessary pipelining for data mining and model selection.
The problem is in the selection of predictors and their selection, there are no problems in models for a long time.
Resampling is done to remove outliers, smooth the sample
1. Resampling an outlier doesn't remove it. There are programmes, and you can do it the kolkhoz way: change everything greater than +/- 0.005 of the corresponding quantile to this value. The statistics change remarkably.
2. Extremely interesting, especially on entropy. I would like to see the result. Correlation is for stationary series, we can forget it.