Machine learning in trading: theory, models, practice and algo-trading - page 2260

New comment
 
fxsaber:

How long it took to ignore the obvious...

It hasn't been ignored for a long time, not since the boxplots article)

 
Maxim Dmitrievsky:

long time ago is not ignored, after the article about boxplots )

Then the obvious evening scalper, holding for many years, should have long ago lit up. The whole Market is full of them. Their sales turnover in the Market is in the millions ($) per year.

 
fxsaber:

Then the obvious evening scalper, holding for many years, should have long ago lit up. The whole Market is full of them. Their sales turnover in the Market is in the millions ($) per year.

I don't have enough power to make an MO for ticks. This would be snapped up, like super profits. And it's more of a positional TS.

I think everything is not so rosy in the market, if you look closely. There are a few relatively normal ones in the top, and that's it.
 
A fun topic is doing a reversal of the TS from the signal mart with the help of the MO. It's actually quite simple.
 
Maxim Dmitrievsky:
is a fun topic - doing a reverse TC with an MO signal mart. It's actually quite simple.

The closed groups/signals of Market Products don't exist out of thin air.

 
fxsaber:

The closed groups/signals of Market Products do not exist out of thin air.

But the reverse is more difficult. With MO on logic. Almost impossible, you can only approximate

MO can only be broken by another stronger MO
 
  1. The Market Product of interest is raced in the Tester - an automatic run of a large number of input vectors.
  2. For each one we took from tst-file (format open) deals and input vector.
  3. And then we fill this DB with a lot of other "indicators" in the points of deals.
  4. In the first half of the database we teach, and in the second half we check.
All four points may be completely automated if desired.


I doubt that some MO would be able to reengineer such a TS: it looks for the most similar to the current sections in the past. And if there is statistically a preponderance of further movement in some direction - that is where we signal.

If before that for simplicity the search of similar segments is performed not on the price series, but on the transformed price series, for example: Entries or bars are replaced by the binary logic: up(0)/down(1). Then the problem of reengineering for MO becomes quite complicated.

 
fxsaber:
  1. The Market Product of interest is raced in the Tester - an automatic run of a large number of input vectors.
  2. For each one we took from tst-file (format open) deals and input vector.
  3. And then we fill this DB with a lot of other "indicators" in the points of deals.
  4. In the first half of the database we teach, and in the second half we check.
All four points can be fully automated if you want.

Yes. In Python it's a fairy tale. But for the ticks it needs a lot of code optimizations, because it is slow. The main brake is training the classifier, all the rest are rubbish. But this slowdown can only be avoided by powerful hardware, all models are in C anyway.

Array handling - also with C speed. Small overheads in the form of python fs and that's all
 
Maxim Dmitrievsky:

because it's slow.

For there is no need to conduct frontal attacks. The advantage is not for those who can create complex models, but for those who can hammer very fast: orders of magnitude faster than the rest.

 
fxsaber:

For there is no need to conduct frontal attacks. The advantage is not who can create complex models, but who can hammer very fast: orders of magnitude faster than the rest.

Huh... instead of a thousand hammers one good modern model

complex models are a huge advantage, but to get to them you need to spend a couple of years of your life, which has already been successfully spent

1...225322542255225622572258225922602261226222632264226522662267...3494
New comment
Reason: