Quantitatively predictive model building has proven itself in physics and could enrich political science, which is presently addicted to postdictive approaches. As a specific example, I address the dependence of electoral volatility on the number of parties in both ways: testing falsifiable predictive models based on logical considerations, and theory-free postdictive regression. The pressure to apply regression to any and all problems harms political science by omitting one-half of the scientific method. Most numbers published in political science are dead on arrival: once printed, they are never used for anything. We can do better than that. It is not a matter of starting from scratch but widening the tool kit so as to bring to completion research that presently all too often stops just short of the pay-off stage.