Thoughts on:

"Transform-Based Similarity Methods For Sequence Mining",
Tara Halwes,
Department of Computer Science, Worcester Polytechnic Institute, 2000

This got me thinking about events. Not events as the paper describes, 
but rather events that correlate to some specific time period in a time
series attribute. Looking for an event, such as a numeric metric's
value decreasing suddenly, and then looking for events that occur during
the same time in other time series attributes of the same instance.
The result could be a rule like: If events A, B and C happen in
attributes 1 and 2, event D will happen in attribute 3, or some such 
thing. The likelihood of an event based on single valued attributes 
might be found. One could expand the system by simply adding a new type 
of event and a method of identifying it.
      
Much of the reference material has been concerned with identifying
sequences that are similar to a query sequence or sub-sequence. That
idea is present here as well (similarity transform). We could
identify associations between sequences
in an instance and then compare these associations to those found in
instances containing similar sequences. This seems a little unclear...
the context is that we could very well mine for associations inside
each instance, treating it as our whole data set, with each time period
being a separate instance in this context, for example. Once 
associations were found, we could go back to our real data set and easily
search for sequences that are similar to those used in the association
rules we have already found. Sort of like finding a rule from mining
one instance then explicitly looking for support of that rule in similar
instances.
      
The difficulty would be in pre-defining sub sequences to search for.