|
PureBytes Links
Trading Reference Links
|
Thank you,
Using the current data the models have been fairly robust. However, the Neural
Generator spits out models that assigns different "weights" to each variable
that it decides to use.
So for example, I may have a strong model that does not "leak" because the
weight of the O.I. variables are slight by comparison to other variables. I do
have three models that leak and I could not "hit" on where the leak was. When
I saw your post I returned to the variables list and sure enough, those three
models have a stronger weight assigned to Open Interest.
In the event I am not clear about what a "leak" is, the example is that the
model would act based on data provided on Monday for Monday and then on
Tuesday would change that Monday action because the data for Monday had
changed. So one might take a positional Long on Monday only to discover on
Tuesday evening that the "proper" trade was a Short.
I think, insofar as the more robust models underweight the O.I. numbers I
shall just discard the variables from the systems. I do imagine the net effect
will be very negative. It is not so much that "absolute" accuracy is necessary
but at market turns when the Signal is on the cusp of a change then a simple
error in a heavily weighted variable will tip the scales just enough to
implode the model.
--Pat Slevin
THE DOCTOR wrote:
>
Patrick,
If your system requires absolute accuracy than open interest is simply the
wrong variable to use. There is simply too much "noise' in the open interest
numbers.They are a decent representation over time, but they are an exact
measure. More importantly how has your modeling, using the current data worked?
There are all kinds of problems with open interest including the customer
double counting so they should never be assumed to a real precise measure.
Sorry...
>
> Patrick Slevin wrote:
>
> > Could you please clarify something? I use EOD O.I. numbers for Neural Net
> > modeling and I wonder, as the result of your post, whether I might be
> > inadvertently instilling "leaks".
> >
> > That is, when you say that the "aggregate" may change, would that mean:
> >
> > If Monday was originally reported as having 2,000 Open OEX Puts, for example,
> > but then on Tuesday Close it was changed to 2,500.
> >
> > A: Would the Monday Close be "back adjusted" to 2500
> >
> > or
> >
> > B: Would Monday's Close remain the same for reporting purposes and Tuesday
> > would be "adjusted" to reflect the missing 500 contracts.
> >
> > This is important for me to know, because if a Neural model encounters changes
> > in past history the Neural signal may change. That is to say, if the model
> > "thinks" there are 2,000 puts open on Monday it's signal may be to Sell. If I
> > tell it on Tuesday that Monday was really 2,500 then it may go back and change
> > that sell signal to a Buy.
> >
> > Obviously, if the O.I is back adjusted then I must discard it as a variable
> > for Net modeling.
> >
> > Thanks in advance,
> >
> > --Pat Slevin
> >
|