|
PureBytes Links
Trading Reference Links
|
Gabe Hanover wrote:
> There is a philosophical argument which I think
> is applicable here: You can not predict A with A.
This argument is way too broad to have clear application.
In fact, there's a whole class of time series prediction tools that
depend on lagged values to predict future values, namely, autoregressive
models (AR, ARMA, ARIMA, etc., etc.)
In brief, anywhere you have a time series in which there are a variety
of short-term adjustments and feedback mechanisms, you might very well
be able to predict A with A[n].
In trader-land you might often hear "The trend is your friend," "A trend
in motion ...", etc., which implies that past values _are_ significant.
I'm not saying that these AR models work in the markets, 'cos I haven't
done the work so I don't know.
+++++++++++++++++
Allan Kaminsky wrote:
> It should also be observed that restricting
> the solution to multi-linear regression may
> yield an inferior solution to using non-linear,
> non-parametric multiple regression.
Sounds neat, but what exactly is NLNPMR?
By my very fuzzy understanding, regression can be made non-linear by
various transformations of the dependent and/ or independent variables.
Or perhaps you're referring to neural nets?
As for non-parametric, not familiar with that in this particular
context.
Oh, boy, a new statistical toy ...
Cheers,
Cab
|