"Predictive Analytics" is the "next generation of data-mining" according to Forrester Research. The promise of significant bottomline benefits is making it big business with most enterprise software vendors participating. I believe that complexity sciences modeling has a generally unrecognized potential to play an important role in the development of this capability.
Predictive Analytics tools generally are combined with "Business Information Systems" to analyze historical data, identify patterns of behavior, and project those patterns into the future to improve real-time or near real-time decision making. The underlying analytic approaches typically rely on statistical models of varying complexity, discrete choice models, or some machine learning approach such as neural net modeling.
These techniques are not as effective in situations that are characterized by a high degree of "dynamic complexity", ie: in situations where time or system feedback play a big role, because these modeling technologies do not handle dynamics very well (see "Let's Get Statistical"). Nor will they be useful in situations where the underlying system structure is changing because because historical behavior patterns will change.
I think of this as the difference between driving by "looking in the rear-view mirror" and driving by "looking out the windshield". If the road is straight then it is indeed possible to navigate by the rear-view mirror. But if there are any curves watch out!
I've tried to summarize some of these ideas in this picture. A complexity sciences model, built using a mix of agent based modeling techniques and system dynamics, is based on rich historical data and on the structure of the relevant system. To produce a prediction of the future the model must capture the decision rules in play at the current time and through-out the time horizon of interest. These decision rules can include the effects of traditional predictive analytics if they are active. The model can produce a set of future behaviors and, if desired, can be used to explore what an optimal set of decision rules would be.
