The Probable vs. The Possible: Which Insights Are More Valuable?

We do simulations so we can visually explore possible outcomes. But when we want to get valuable insights, should we focus on the possible or the probable?

As a case study, we can use the wide variety of predictive models that well-known media outlets have been promoting for the U.S. Presidential race. We created our own--the Aggregators Ensemble--that tended to show a higher likelihood, at any point in time, of candidate Hillary Clinton winning the election than most other predictive models.

Our Aggregators Ensemble Presidential prediction model as of Tuesday, October 18. (For the full interactive go to

Why is our model different from other models? One way is in its emphasis on the probable over the possible.

Let's take a look at one of the reddest states and one of the bluest. The best poll aggregators out there will place Wyoming as >99% and Hawaii as >99%. Hawaii is probably going to remain a blue and Wyoming red. This is why as soon as polls close for a state on election day, we hear news outlets immediately announce "with 1% of districts reporting, we call this state for..."

We want realistic insights. To get those, the art fills in the gaps left by the science. 

Now here's where running simulations goes from academic to actually insightful. If we put Hawaii as 99% going blue, then when we run 100 simulations, Hawaii will be red 1% of the time. Take another >99% blue state like California. Should we really run simulations where there is an academic possibility of it becoming a red state for the first time since 1988?

We want realistic insights. To get those, the art fills in the gaps left by the science. We make the call to place probability for some of those states at 100% and then run our 20.000 simulations.

This is almost certainly not being done by some of our better known prognosticating colleagues. So when we have election results with an over 98% probability, that comes from our confidence states like Wyoming will be red and California will be blue. As a result, our model has less volatility, because it discards outcomes that are mathematically possible, but provide no real-world insight.