One of the major questions that came out of the last week’s US election was why various polls and forecasts were unable to predict the outcome. For clarity, there are three general kinds of predictions when we are talking about elections: opinion polls, exit polls, and forecasts.
Opinion polls are run by a variety of organizations in the run up to an election. Exit polls are run by Edison Research and the National Election Pool and occur during voting. Forecasts are mathematical models various groups assemble based on differing datasets to attempt predicting the outcome. In this election, as in previous elections, the exit polls were correct.
Opinion polls varied in correctness, and forecasts were generally incorrect. And some political campaign polls that were poorly constructed or biased ended up being more accurate, albeit as just a coincidence.
Why can’t something as time-bound and important be accurately forecasted? Here is the simple explanation and a golden rule of predictive analytics: you cannot forecast something that has never happened. All models and forecasts require previous data, and the more regular and frequent that data, the better the forecast. Elections fail on both counts. Political elections are very infrequent, compared to things like e-commerce transactions. Elections are also highly irregular. Why are election events so strange?
Elections may happen regularly in America, but what is being voted on varies wildly from election to election. Each candidate is different - even if the same candidate is running for re-election.
Imagine trying to forecast how two movies will perform against each other, but it is a new and different set of movies every time. A romantic com vs a science fiction flick. A documentary vs a horror film. It is never the same politician lineup, so it is very difficult to forecast. Now imagine trying to do the same thing, but the movie theater moves around. One year it is in New York City, the next year it is in Peoria.
Will Captain America play better in one location than another? What about the newest Fifty Shades of Grey movie or spinoff? That is what has been happening in American elections as the population changes.
The America that voted in politics and pop culture last week is not the same America that voted in 2020, or the same America that voted in 2016. Add onto that all the hidden inputs that normally mess with models, from differing marketing tactics to foreign interference to social networks, and suddenly we have an insurmountable amount of relevant data we would need to forecast an election in 2023 or 2024. And then add a pandemic. Is it any wonder forecasts and models failed to predict the outcome?
Here is the lesson we must learn from politicians as marketers: you can’t forecast what has never happened. If you are in the throes of 2023 planning right now, you are probably looking back at past data and trying to project it forward. Beware! The world you operate in right now is very different than the world in January of 2020 - completely different due to global events. This is a new world, a new operating environment where the customer does fundamentally different things than they did 12 months ago.
Forecasts and models built on last year’s data almost certainly will be wrong. What should we do instead? Rely heavily on near real time data. Look at trends that are recent, rather than pre-pandemic. If we are trying to forecast something new - a product or service we haven’t ever launched, the tenure of a new executive, etc. - we should expect our forecasts to potentially be radically wrong.
And be doubly cautious of vendors promising they can reliably forecast the future. No one can reliably do that now for any substantial time horizon.
It is a brave new world for all of us, and nothing is guaranteed now. Focus not on what could happen, but what is happening right now. We still don't know exactly what is going to happen with a new presidential inauguration, the upcoming 2023 and 2024 races, or overseas elections either. I guess it is politics as usual: expect the unexpected. Time will tell!
