Machine learning in trading: theory, models, practice and algo-trading - page 3669
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
Theoretical question. Suppose there is a certain trending asset. Due to the trendiness, it makes sense to make a trailing stop exit.
Does it mean that it makes sense to build the model only as a regression? That is, use the distance travelled by the price before the exit as the exit Y.
And reduction to classification (by fixing the exit level) will be some kind of distortion of the TS essence?
That is why I have been training only regression models for 2-3 years.
And that's why amateurs of classification, achieving an error of 20%, lose money.
Will be a distortion. Classification will remember how many times you won or lost, but not how much in money/points.
That is why I have been training only regression models for 2-3 years.
And that's also why amateurs of classification, achieving 20% error, drain in money.
It depends on WHAT to classify. Formulating a "teacher" is not an easy task. For example, a "teacher" is just an increment, or a teacher is an increment of more than some number of points. But even this is not enough for a quality "teacher".
And in regression another problem, which is the same with classification, is the forecast error in points, which makes the direction of entry uncertain.
Will be a distortion. Classification will remember how many times you won or lost, but not how much in money/points.
That is why I have been training only regression models for 2-3 years.
And that's also why amateurs of classification, achieving 20% error, drain in money.
Then technical issues arise. For example, we do regression on trees. If the outputs are ordinary increments on a candle, it is logical to use LS for splits, because the increments have a shape more or less close to Gaussian. If we make regression for exits - maximum price movements before triggering the trailing stop, their distribution is not Gaussian at all (rather exponential). Imho, this requires changing the split rule when building the tree from LS to an adequate one.
LS is what?
I don't analyse the shape of distributions. Just testing different ideas/variants. So far I haven't found anything stable.
At least I can immediately see that it's 50/50 in money, not 20/80 in pieces with unknown money filling in each such piece.
LS is what?
I don't analyse the shape of the distributions. Just testing different ideas/variants. Haven't found anything stable yet.
At least I can immediately see that it's 50/50 in money, not 20/80 in pieces.
LS - Least Squares, the usual minimum squares that usually defaults as a metric for splits in regression.
Well, the question is not about practice, but more about theory. There will probably be one or two very long trends that will mess things up when using LS. As outliers, although they are not outliers in the essence of a trending TS.
Perhaps the correct question is: Is it possible to capture rare but long movements with the MO model at all?
I wanted to write something like this, but then I fell asleep and dreamed that I had already written it. As if the usual MO logic breaks down for trending TSs.
This is very, very unfortunate.
Perhaps the correct question is: Is it possible to catch rare but long movements with the MO model at all?
It's really, really unfortunate.