[R-SIG-Finance] Question about Linear Models, Machine Learning

[Alex Grund]

Hi @all,

I would like to conduct a research about the impact of macroeconomic
variables to stock market returns, similar to APT.

APT suggests a linear model

r = a + b1 F1 + b2 F2 + ... + epsilon

where r is the return, and F1,... are factors, the rest follows a linear model.

However, when putting macroeconomic factors in such a model, chances
are high, that the model is over- or underfitted, i.e. having too less
or too much factors in it.
The next thing is that each factor has to be transformed, for example:
Factor 1 may be the GDP of USA, but in order to come into the model,
it needs to be diff(log())ed and lagged by 1 or 4 or whatever.

In machine learning there are many suggestions for unsupervised learning.

So, imagine, I have a bunch of indicators and "next month returns" of
some stock market index as dependent variable. I want to fit a linear
model, not knowing the causalities behind it. And I want R to estimate
the transformations and lags to find the model itself, not only its
coefficients. Which machine learning technique is for that? And is
there an easy to use introduction to it and maybe even some R modules?

What should I search for?


Thanks for any hints
--A