Estimating electability from betting market data


First, thank you to all the bokeh developers for creating such a great package, being so helpful in answering questions, and for generally being just the nicest people.

I made a simple bokeh server application which shows time series of an estimate of ‘electability’ of major candidates in the 2020 US presidential election. Calculation details are in the text below the charts:

Hope you enjoy it! Any questions let me know.



This is both an interesting topic and an easy-to-follow, concise presentation of the data. The well-written summary of the methodology behind the data and the limitations are also quite helpful.

One question regarding the lower graph on the hosted URL. The lower plot Probability of winning presidency if chosen as nominee has the sum of probabilities > 1.0 in places. I can see conceptually how the sum might < 1.0 if there were candidates not listed, for example.

Focusing on the data from 2020 MAY onward, what accounts for the sum of winning probabilities exceeding 1.0 beyond the precision of the data reported? (This seems to show up consistently in the curves and the hover information)

How is this method similar to 538 models? Or do they not describe their methodologies in enough detail to know?


Thank you for the nice words and good questions.

The conditional probabilities on the lower chart won’t sum to 1 because they correspond to different states of the world (a state where Bernie wins the nomination, a state where Biden wins, etc). As a thought experiment, assume for simplicity that Trump is definitely going to be the GOP nominee, that he is wildly unpopular, and that any Democratic candidate would definitely beat him. Then all the Democratic candidates’ conditional probabilities would be 1 because their probability to win the nomination would equal their probability to win the general election.

The post-May numbers are a little harder to explain but I’ll make a guess. Currently the presidential election betting market is saying Biden has a 60% chance to win in November and Trump has a 40% chance (you can see this market on predictit website). This means the conditional probabilities will only sum to 1 if each candidate has a 100% chance of winning their nomination, but in fact the nomination markets are saying that both have about a 90% chance. It seems that the presidential election market is ignoring the low-probability events that one or both men don’t actually get their party’s nomination.

I think the 538 models do take this betting market data into account, but they also use a lot of polling data and have different weighting schemes depending on the quality and size of the polls etc.

Hope this helps!

1 Like