Wednesday, August 20, 2008

Prediction, system dynamics, and Future-Fusion

Recently, I made the claim that we're better off focusing on adapting to the present than predicting the future. I've made similar claims in the past, too. I've even given one example in which predictions serve a useful purpose.

That's all a bit simplistic, of course. Even system dynamicists could be said to predict the future in a way: we show behavior over time we feel is more likely to occur (although we may warn people away from point predictions based on a behavior over time graph). In other words, I might suggest that your current policies could produce a boom and bust effect in your business, but I wouldn't want you to draw the conclusion that your business will grow another 172.3% by June 15, 2009 before taking a tumble that afternoon.

Because we all would like to know the future, I've experimented with blending system dynamics and Bayesian analysis to quantify the probability of a particular behavior pattern, for example. Of course, that probability is conditioned on both the historical data and the model being correct, which is a loophole big enough for a good-sized locomotive to run through: models are always incorrect. Still, I think this approach may give more useful insight in certain cases.

Now Kshanti Greene of Stottler Henke Assocates, Inc. has shown me a Bayesian tool they've developed called Future-Fusion, and I've been exploring it a bit. They are using Bayesian networks and the power of groups to get a better handle on what the future holds. Much as Data360 looks at the past, Future-Fusion attempts to look at the future. As of this writing, they've created four test areas which you can explore: the 2008 US presidential election, the Iraq war, corporate strategy, and energy. Try it out: learn how to use the system, see current predictions, and add your own (I think you only have to create a free account if you want to add your own predictions). Perhaps you'll learn something, and perhaps they will, too.

Kshanti has pointed out a recent addition to Future-Fusion that may intrigue some of you: time. They've enhanced their technology to allow limited dynamic execution of a network model, which begins to narrow the gap between Bayesian networks and system dynamics from the Bayesian network side, much as what I've tried has narrowed it from the system dynamics side. To try that out, go to the energy model, select a prediction (e.g., "Reduced SUV sales"), click "view graph," note the numbers, and then click "Next Time Step."

I think this is all still experimental in many ways, but it's a good opportunity to learn a bit about this technology by trying it out on real-life issues. I'll be curious what you discover.

Labels: , , , , , , , , , , , ,

2 Comments:

Blogger Innovative Thinking said...

I really liked your idea of using conditional probability for prediction purpose.

I am also doing the same with my models and find it to be useful.

However, I find that it is easier to apply on a balanced system and a little difficult to apply on a reinforcing system.

Any help?
regards,
dibyendu

02 November, 2008 06:09  
Blogger Bill Harris said...

dibyendu, I've tried that on systems with interacting balancing and reinforcing loops, and it seemed to work okay. When doing that, I've been using MCSim. You can find some links about MCSim on my Web site, and you can find the latest software on Savannah.

02 November, 2008 13:48  

Post a Comment

<< Home