As of the moment I write this, we don’t know who will be the next President and probably won’t for a few days. What we do know is the projection sites and the state polling were off base again and the election was much closer than anticipated.
For example, using the 538 dataset, WI was forecast as a 9% Biden win and MI at 8% but each will likely be <1%. Looking at states Trump was expected to win, TX and OH were each forecast as a 1% Trump win and ended up +6% and +8% respectively. Even the blowout states generally overestimated Biden by 5%.
While it’s a little early to say exactly what the polls missed, I have a feeling I can figure it out. Loyal readers will recall I like to talk about nonstationary data. This is the concept that relationships change over time so building a model off of history may not represent the future well.
Nonstationarity is a big issue in financial markets which is why every commercial for an investment product reminds you about past returns not being predictive of the future.
However, it also applies to political polls. Pollsters don’t call 50 Republicans and 50 Democrats and tally the results. They know response rates are low so they get as many valid responses as they can.
Then, they take this raw data and adjust it for demographic differences. So if they get more Democrats than Republicans in their sample, they will adjust down the raw data to make an estimate of the results they would have gotten if they had say equal R & D responses. They make similar adjustments to the raw tallies for age, income, education, etc.
The problem with this approach is it assumes the pollster knows what the right “steady state” is to adjust towards. Because the path of least resistance is to assume the mix of voters will be the same as in the past, the pollsters anchor their forecasts to the makeup of prior electorates.
In other words, the polls will do a good job of capturing how many Obama voters switched to Trump last time or how many Trump 16 voters will vote for him again this time.
Missing the Turnout Impact
They will do poorly at capturing changes in who votes. Given all reports indicated this would be a massive turnout year, the prior elections were poor bases for the actual outcome in 2020. The change in turnout would potentially dwarf the number of people who voted for Trump last time, but didn’t this time.
Furthermore, we know Trump’s re-election strategy was based on turnout of his base not expanding it. Additionally, the Democrats had a very strong get out the vote operation across social media and sports & entertainment platforms which seemed to be successful based on lengthy early voting lines in Democratic leaning areas.
There was every reason to believe the enhanced turnout would decrease the accuracy of polling. In fact, I had this exact discussion with people before the election.
I suggested the polls would be way off because they’re assuming the wrong makeup of the electorate, though it wasn’t clear in which direction, especially given the prevalence of early and mail voting.
Now that we have data, we can see the polls underestimated the Trump turnout more than the Biden turnout and thus we have ended up with a much closer election than expected.
Lessons For Actuaries
How is it that I, as an “outsider”, was generally better at predicting the direction of most company’s reserves than the actual actuaries who had much better data?
Because I understood loss trends weren’t always stationary. Most actuaries rely too much on their internal history and underweight the impact from the external environment.
For instance, it was obvious lawyers found a new way to attack D&O policies several years ago, but it took forever for that to make its way into reserve estimates.
Or take commercial auto where actuaries have been behind the curve for close to a decade now even though there was clearly a massive change in severity that you could trace with a little bit of persistent googling.
Actuaries do a terrible job at recognizing change in trend and continue to extrapolate what they view as “credible” data (credible because there is lots of it, not because it is predictive) and thus compound their errors.
Part of being good at using data is knowing when to rely on it and when to question it. I have talked about this several times now. Unfortunately, not many people are good at it. Better to go off and square another triangle or dial 1000 more “likely voters”.
Aside: The Electoral College
One last thought: many people, when we have these tight elections, question whether the electoral college is the best way to elect a President. It probably isn’t the best way, but neither is it the worst way. Thus, in the absence of clear evidence that another method is substantially better, we should stick with the status quo.
Let me address a few of the concerns. First, isn’t it disenfranchising to voters in bigger solid states to let smaller swing states decide the country’s fate?
If everything were done in a vacuum, then perhaps. However, campaign spending is considered speech. There is no restriction on people who live in one state spending on an election in another. We have seen Bloomberg try to buy Florida for Biden. This is his right.
Every time a citizens donates to one of the candidates, they realize that money is likely going to be spent in a swing state. So everyone does have the opportunity to influence the swing state voting, though it is indirectly (through $) rather than directly (by voting).
Second, why not just let the popular vote decide? While we can debate the merits of letting the big states be the new swing states, frankly, that isn’t the right debate.
We are a representative democracy, not an absolute one. If we switched to a popular vote for elections, wouldn’t that suggest all decisions should be made by popular vote? So disband Congress, disband all the reguatory agencies, and let people vote on every policy.
If we want to be a representative democracy, then we should continue to elect representatives to pick the President. We can debate whether the electors should be say the members of the House rather than unknown electors, but going to a straight vote is incompatible with a representative democracy.
Finally, most people who complain about the Electoral College only do so when their side loses. When they benefit, they don’t mind. This is not to say there isn’t room for new ideas, but the ideas should be supported by data, not partisanship.
Swing states do change over time and, for most of our history, competing in the swing states forces candidates towards the middle to expand their coalition. This generally ensures that candidates do represent the median view of the country which is arguably the goal of elections.