Machine learning in trading: theory, models, practice and algo-trading - page 501
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
You understood nothing from my post. Nothing at all.
And it is a pity for the branch because of your presence.
Yeah, you then define your comrades in the shop, who claim to be able to extrapolate :))
In fact, the question raised is interesting.
In the case of regression y = f(x), the forest really cannot "extrapolate" beyond where there are no training points, while MLP can, although it does not do this either, in the style of polynomial regression at most, which is "prettier" on the constant's chart from trees.
But Sansanych and Leha are also right in some respects, because regression y = f(x) is not what we traders need, because if x is time from Birth of Christ, we are not interested in such relations, and in space of other features, points from future time will not necessarily be outside the learning sample from our indicator features.
Why a forest of trees gives a constant is clear to me,linear regression is also obvious, but why MLP gives a pseudo-polynomial regression I "don't see" as transparent. we need to figure it out...
Well a simple example, if you predict prices and train the scaffold on an incomplete price interval, then in case you need to predict a price outside the training interval then it won't predict it. It is clear that these situations are specific and you need to train correctly and then you can avoid such problems in most cases
your picture is exactly what i'm talking about
Tell me HOW?
Paint ))
No problem at all, because there is no linear time in the signs, "outside the training interval" only very rare spikes, something that should not be, fucking black swans :)
ah, i thought it depended on the target interval... i must have been inattentive
Forum on trading, automated trading systems and trading strategies testing.
Table with quotes for 1000 of the most traded stocks. Data for the last year.
fxsaber, 2017.10.04 14:19
"Regularly rebalance the portafloor and all will be well" is what so many people say, until they slip into their counting system SB (random walk) instead of the original price series.
The vast majority after SB see exactly the same picture as with real data. But to draw the same conclusion by saying the phrase about rebalancing is no longer possible, because it is SB. And "will be good" is ruled out.
The question, as I see it, is the key one: how much does the real price data differ from the SB? If I understand correctly, the greater the difference, the more opportunities to squeeze out a profit. And vice versa, up to "no difference - no profit".
why MLP gives pseudo-polynomial regression I "don't see" as transparent. need to figure it out...
Figure out why 1 neuron does linear regression, how it happens at low level, then you can easily understand how non-linear variants are obtained in multineuron and multilayer sets. And Random Forest as well as Knn and other kernel tools do only local interpolation, which may be extrapolation in other projections, not the point, but RF does not build(sweep) functions in wide range, but multilayer perseptron does.
The question, as I see it, is key: how much does the real price data differ from the SB? If I understand correctly, the greater the difference, the more opportunities to squeeze out a profit. And vice versa, up to "no difference - no profit".
They are different, but very insignificantly, so you cannot see them by eye. Besides, most of the differences are already included into trade costs.
I would like to ask the respected participants of this branch on this subject
The question, as I see it, is key: how much does the real price data differ from the SB? If I understand correctly, the greater the difference, the more opportunities to squeeze out a profit. And vice versa, up to "no difference - no profit".
The differences consist in the presence of a relatively large emission in the price series and the presence of a weak "seasonality" - hours/trading sessions, days of the week, etc.
Well, also "fat" tails.
The differences consist of the presence of a relatively large emission in the price series and the presence of a weakly pronounced "seasonality" - hours/trading sessions, days of the week, etc.
It is possible to glue together several SBs with different intervals to get the same effect.
Well, there are also "fat" tails.
Take a "thick-tailed" SB.
You can glue several SBs at different intervals to get the same effect.
Take a "thick-tailed" SB.
It is possible to fit SBs to the signs of the price series, but after all these transformations it will no longer be an SB.
What is the point?