GEN: phase behavior in price action

[Kevin Morgan <kmorgan@xxxxxxxxxx>]

I've got a notion about market action I'd like some feedback on.

In computer programs there are memory access patterns that occur over the
course of execution of the program.  Specifically, a program moves through
"phases" of behavior, and in each phase, a particular pattern of memory
accesses occurs for a period of time.  A diagram of this can be built by
putting graphing memory locations (by pages) long the Y axis, and time at
some unit of granularity on the X axis.  Black in the areas of memory
being touched across time.  You'll see fairly consistent patterns, then
suddenly, a change to a new pattern, etcetera.

Market price/time graphs exhibit the same thing.  A pattern
emerges, goes for awhile, then changes, goes for awhile,  changes, etcetera.

Might it be possible to:

  - build a canonical set of "phase behavior" patterns for a given market,
    or even for "all" markets;

  - have an algorithm for "early detection" of probable phase change and what
    the new phase type is;

  - select the trading algorithm best suited to the likely phase type;

  - execute that trading algorithm, and make real time adjustments if
    the phase identification component changes it's mind about what's up.

Obviously, the key to success is early and accurate detection of phase
action, combined with the quality of the trading algorithm against that
type of phase bahavior.

This idea has been rolling around in my head a little.  What do you think?
Does anyone know of published material along this line of market analysis?
Is this a new idea or old hat?

I realize the most basic approach to this is "trending" versus "non-trending"
market assessment, and adaptive trading algorithms based on this assessment;
this indeed is "old hat", but...difficult (in my system building experience)!  
What I'm suggesting here is a generalization of this approach.

A possible way to proceed would be to have a canonical set of phase behaviors
which are parameterized, and the parameters are identified based on specific 
market behavior in the past, with more emphasis on recent market behavior.  
Another would be to just do all phase analysis/pattern construction "on the 
fly" based on specific market data (although I guess there'd still have to be
some "archtypes" of phase action that the analysis s/w is looking for, so
maybe this isn't really any different).

All ideas and feedback are welcome, sorry if this is out in left field.

Thanks,
-Kevin