Machine learning in trading: theory, models, practice and algo-trading - page 3142
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
That's a long search, especially if you search for a few seconds.
Yes, long, really not in a hurry, but still long.
The problem is that the predictive power of predictors is only one of the chips. And there are many of them in my preprocessing and each requires a set of statistics to justify it.
I mentioned one more problem above, for which I can't find a solution for a long time.
For a new teacher, no.
I am trying to solve the problem of coarsening of predictor values. It seems to me that a classification error can occur if the predictor value is slightly different from the value on which the model was trained. I once tried to convert all predictors into nominal form with the same teacher, but it did not give any result. However, the number of values of nominal variables was one. Maybe we need several hundred? I am working, but many other interesting questions are in the way.
"
An ideal model should have low bias and low variance. However, in reality, there is a so-called "trade-off" between bias and variance. Increasing the complexity of the model (e.g., adding more parameters) can reduce the bias but increase the variance. Conversely, simplifying the model may reduce variance but increase bias.
"
Hypothetically you can bend the crankshaft to a satisfactory level and sharpen the clearances, but that won't go far because it's no longer "by design"."
An ideal model should have low bias and low variance. However, in reality, there is a so-called "trade-off" between bias and variance. Increasing the complexity of the model (e.g., adding more parameters) can reduce the bias but increase the variance. Conversely, simplifying the model may reduce variance but increase bias.
"
Hypothetically you can bend the crankshaft to a satisfactory level and sharpen the clearances, but that won't go far because it's no longer "by design".For some reason, the third component is often forgotten - unrecoverable error. If it is big enough (and it seems to me that we have a lot of it because of the proximity of prices to SB), it may be more important than the first two.
In any case, these are very important things that can be put into one question: What is the maximum information that can be extracted from the sample of prices (and other available data) that we have?
For some reason people often forget about the third component - unrecoverable error. If it is large enough (and it seems to me that in our country it is not insignificant because of the proximity of prices to the SB), it may be more important than the first two.
In any case, these are very important things that can be put into one question: What is the maximum information that can be extracted from the sample of prices (and other available data) that we have?
so we shouldn't try to predict all cases, but pull out the ones that are predicted through mo
this is called "heterogeneous tritment effect", which can be compared not to bending a crankshaft, but to finding working parts and discarding non-working ones.
then the attributes of X become contextual and are not "predictors" for Y in the classical sense. That is why in kozul they are called "covariates".
The result will answer your question (depending on what to measure in) about the maximum information. Usually it is measured in ATE or CATE.
One question: What is the maximum amount of information that can be extracted from the available sample of prices (and other available data)?
For some reason people often forget about the third component - unrecoverable error. If it is large enough (and it seems to me that in our country it is not insignificant because of the proximity of prices to the SB), it may be more important than the first two.
In any case, these are very important things that can be put into one question: What is the maximum information that can be extracted from the sample of prices (and other available data) that we have?
That's a question of dissertable research, not of building a robot that mows down dough.
We don't need a maximum of information, we need a sufficient minimum. Therefore, we can limit ourselves to the following:
1. Create a model that gives a classification error of less than 20% at the classification stage. And it should be understood that the "model" includes full preprocessing of predictors, as well as tools for model evaluation.
2. Insert the model into an Expert Advisor that gives at least the same ratio of losing/profitable trades. If the profit factor is above 4, there is one more step left to take.
3. Make sure on the OOS that nothing has changed, and understand the reasons for such stability on the OOS, which lies in preprocessing, not in the model.
And what errors got into 20% - is it interesting?
Wouldn't it be easier to classify errors?
the more errors we find, the better the model will remain. not maximisation of information, but qualitative (pardon the pun).
I think that Snalo needs to set the task properly.
In my opinion, Maxim's option above is very good. As San Sanych rightly noted, it is not the information itself that is important, but how it helps to multiply the deposit).
If for simplicity we consider a trading strategy as a tritent and Y as a profit, then the definition (TS maximising profit expectation) becomes quite banal.
It's a matter of dissertable research, not creating a robot that mows down dough.
We don't need maximum information, we need a sufficient minimum. So we can limit ourselves to the following:
1. Create a model that gives a classification error of less than 20% at the classification stage. And it should be understood that the "model" includes full preprocessing of predictors, as well as tools for model evaluation.
2. Insert the model into an Expert Advisor that gives at least the same ratio of losing/profitable trades. If the profit factor is above 4, there is one more step left to take.
3. Verify on the OOS that nothing has changed, and understand the reasons for such stability on the OOS, which lies in preprocessing, not in the model.
And what errors got into 20% - is it interesting?
One does not interfere with the other in any way. Obviously, the maximum is not only unattainable, but even impossible to calculate, but we can try to estimate it somehow and compare it at least roughly with the spread, for example. I mean a variant of the maximum like the one proposed by Maxim.
One does not interfere with the other in any way. Obviously, the maximum is not only unattainable, but even impossible to calculate. But we can try to estimate it somehow and at least roughly compare it with the spread, for example. I mean a variant of the maximum like the one proposed by Maxim.
Information is in a person's head, but there is no information outside the head.
Everyone has different heads: someone sees trends in a quote, which, by the way, there are many different ones, and someone sees increments. And different people see a lot of other things in the initial quote.
If it is for trends, then the flywheel is more informative than the ZZ, because it rests on the right edge, while the ZZ is not?
Completely vague statement of the problem, which is unnecessary. So maybe there is no such problem?