This week’s events should be interpreted as a word of caution about predictive analytics. Clearly, many models didn’t predict the outcomes of the 2016 election. More importantly, the vast majority of models weren’t predictive. “Models of models” (averages across models) weren’t predictive. Even when models were built on data with high granularity (subnational polls, or polls taken at regular intervals, or polls taken by different houses using a variety of methods).
What’s the upshot? Humility. How much harder is it to get the predictions right when we’re developing policy for new and novel problems?
Don’t believe those who say that big data will solve everything.