Composite Timing Models

["Larry McBride" <lcmcb@xxxxxxxxxxx>]

Here's a question about the development of market timing models I haven't
seen discussed clearly.  It has to do with the development of robust models
from a statistical perspective.

Assume you have three different timing signals you've developed separately
that are based upon reasonably different market parameters or else different
ways of modeling the market.  For example, you may be trying to time the otc
composite for mutual fund trading and you develop three separate timing
signals:
1) an percent swing system based upon and down price action.
2) an indicator based system that totals a stochastic and rsi oscillator,
then trades off the combined oscillator.
3) a relative strength signal that trades based upon the relative strength
of the otc versus some other index.

Now, you develop each of the three signals separately based upon in-sample
data, accepting them if they look reasonable on out of sample data.  Of
course some look better than the others in out of sample observations.  The
ones that look best out of sample are very often not the ones that looked
best in sample.

Now, you build a composite timing model that does something like buy when
any 2 of these signals are long and sell when less than 2 of them are long.
You determine the most effective voting process by looking at the in-sample
data only. Almost invariably, it has looked to me like you would have been
better off trading such a composite in the out of sample data than you would
have been in trying the pick the best individual signal from in sample test
results.  Although this is a heuristic observation, I've seen it enough to
make me think there is something to it.  It has remained apparently true as
well if I use three different time periods instead - in sample, out of
sample & independent verification.

My question is the following, which is really statistical in nature.  Is
such an approach (composites of separate timing signals) likely to be more
robust in the future since the composite model is less dependant upon the
success of any given signal, or is it likely to be less robust since more
parameters (the total of those involved in all signals plus the number of
signal defined in the voting process) are nonetheless involved in the
composite model?

Thoughts anyone?  Thanks for your ideas.
Larry