Machine learning in trading: theory, models, practice and algo-trading - page 3053
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
I tried to write a simple q-table (without this super formula for calculating the action).
It may be a consequence rather than a cause, but it is still a connected phenomenon, and so as long as there is a connection, we can speak of a pattern.
Wejust have to recognise that laws change. And it is when they change - the moment of time - that this problem should be solved.
Laws do not change. We just don't understand all the laws.
In nature, no two snowflakes are exactly alike, but we perceive them as snowflakes and they all look alike to us.
In the financial markets, a pattern is the image of a natural snowflake.
No two patterns are exactly alike in certain parameters as in the case of snowflakes and at the same time patterns exist as a Law.
patterns exist as Law.
if we present all timeframes on a single chart, will we take into account all information about all extremes on the chart?
there are three timeframes on the chart at once m1 . m10, m60
What if we set the task of finding rules a little differently:
the rules themselves need to be different.
if the signs are X1, X2..... Х10
then the rules for wooden ones look like X1>= 0.001 & X2<0.05 .
where 0.001 is just an abstract number/constant in no way tied to the market, just a number for approximation ...
Because of this, when the market changes, all these numbers/constants stop working immediately...
you need rules like X1 >= ( X2*(X5/X7) ) & X3 < (X2^2) * X10.
instead of abstract constants - an adaptive formula,
and constants should be avoided like fire.
they are neither good nor bad, they are adaptive and not constant.
they're neither good nor bad, they're adaptive and not constant.
Once again, I'm not afraid to bore you.
The problem is not the rules or the models, the problem is the predictive power of the predictor for the teacher, which (predictive power) varies. With your approach, you can get into random "goodness of fit", and you need a numerical measure of the variability of predictive ability. You have some analogue of "impotence" for predictor in models - we take from what we have, and if everything is rubbish, we take from rubbish. To me, filtering is a dead end, because the algorithm forms rules on what it is given: if you give candy, you get chocolate, and if you give rubbish, you get another portion of g...rubbish.
the relationship between attributes is just as likely to go out of range. Exactly the same abstraction.
But this is not the point, but the proposed approach, which makes some sense.
I keep waiting for normal thoughts from the forum on how to improve such a thing, because my head rarely gets fresh ideas until I read a couple more books on statistics and IO.
I specifically took the OOS where the market changed. The study was on a falling one and the OOS on a rising one.
How many months/years on the chart?
The OOS is about 75000 pts / 500 deals = 140 pts per deal. That's pretty good, you can put that in a trade.
How many months/years on the chart?