Forecasters Predict the Election (almost)
Editor’s note: Forecast is a social prediction market: an exchange where participants use points to trade on the likelihood of future events. Community members pose questions about the future, make predictions, and contribute to discussions leveraging their collective knowledge. Download the app to join the community.
This post was originally shared on Substack on November 6.
It’s been a week since election day and the dust has mostly settled. Barring some low probability events, it’s likely that the next president of the United States will be Joe Biden. So you’re probably wondering…
What happened on Forecast???
The short answer: In the lead up to the election Forecasters correctly predicted the projected presidential election outcome in all 48 of the states that have been called by the Associate Press (Forecast’s source for settling election outcome questions). Our pre-Election Day map showed a tight race — closer than 60/40 — in the remaining two states, Georgia and North Carolina, but had the more likely outcomes flipped from where they now appear to be netting out. Not bad, forecasters!
Here’s the pre-Election Day map (darker/brighter colors represent higher likelihoods):
For reference, 538 got 47 out of the current 48 right (they missed Florida) and PredictIt also got all 48.
What happened on election day?
The week of the election saw our biggest spike ever in forecasts, forecasters, and, reasons. Overall, just under 500 forecasters made around 10,000 forecasts — double the previous weeks’ totals. Here’s a gif of how the election map changed (you can also see a playable version here). Keep an eye on Florida.
As you can see, there was some volatility in the week leading up to Election Day and then a ton of volatility on Election Day itself. Looking at Florida in particular, you’ll notice that the predicted outcome swayed repeatedly from Biden to Trump and back, before settling at a high likelihood of a Trump win shortly before the AP made its call.
The overall Presidential Election forecast was also highly volatile on Election Day (playable version here).
As we’ve noted before, the big swings are aligned with moments where Forecasters add new reasons (the in-app action where they explain their forecasts). These reasons are in turn timed with breaking news events. As the community assimilated new information (exit polls, ballot count announcements), they continually updated their priors.
What did we learn
The forecasts don’t end when the polls close. Initially, we decided to suspend trading on election outcome questions when polls closed for each state. Our thinking: there’s no new info after the polls close, so why would people want to keep forecasting? Turns out, that was a bit short sighted. As new information in the form of election results started to come in across states and precincts, forecasters let us know — loudly — that they wanted to update their predictions.
10 minutes is a lifetime when election results are rolling in. Forecasts made in the iOS app show up on the Forecast website on a 10 minute time delay. No big deal on a normal day, but with hundreds of trades are coming in over the course of minutes, the difference between overall forecasts the app and website got pretty stark. Live and learn!
A final note
Forecast can be useful — and reasonably accurate — in predicting events like elections. But we think their real value lies in how they help forecasters process information about the news. Forecasters work together to understand and incorporate complex, often conflicting information into an authoritative picture of what’s happen now and what that means for the future. We think that’s pretty exciting.
All the data used to make the visualizations above is available in the Forecast Data Stream.
Please get in touch if you have feedback or otherwise want to talk about Forecast. As always, we invite you to join the community and start forecasting!
