Track Record - Election Betting Odds

By Maxim Lott

Updated Dec 11, 2022

ElectionBettingOdds.com has tracked some 807 different candidate chances across dozens of races between 2016 and December 2022.

How accurate has this site been?

When a candidate favored by bettors loses,
it's easy to fall into the trap of thinking the bettors were "wrong".
But remember that when bettors give someone, for example, a 75% chance
to win -- that also means the bettors give the person a 25% chance to lose.

The following graph shows how closely the betting tracked reality.

The blue line shows what predictions would
look like if they were perfect. The orange line shows how often bettors'
predictions within different ranges (e.g., between 60-80%) actually
came to pass. Where the orange line differs from the blue line, the
predictions were off somewhat:

We can see that, using these prediction
ranges spanning 20 percentage points, ElectionBettingOdds.com was
extremely accurate. For example, out of all candidates who were given
between a 20 and 40% chance of winning at 8am on election day, 27.6% of
them actually won.

We do also see some systematic "underdog"
bias at the edges. For example, for candidates in the 0% to 20% range,
bettors gave an average prediction of a 5.4% chance. Yet, only 1.7% of
those candidates actually ended up winning. The inverse occurs for
highly-favored candidates, who should have been given even higher odds.
This bias has been in part caused by regulatory trading limits on
bettors at PredictIt, which prevents savvy traders on that market from
making much money by correcting such odds.

The graph below holds predictions to a more precise standard -- 10% intervals, with smaller sample sizes.

We can also see that the accuracy of the
odds has been improving over time, as the number of races tracked has
risen and reduced randomness. For comparison, here was the same graph
after our 2018 update, where some 450 candidate chances had been
tracked:

How does all the above compare to other
outlets? For a detailed comparison to the predictions of the best
forecaster (Nate Silver's 538) see here.
That analysis goes all the way from 2016-2022, and does a precise
apples-to-apples comparison between the two prediction methods.
Considering "experts" and models more broadly, several studies suggest
that betting beats listening only to polls. See this one in Public Opinion Quarterly. Also, a study by the author of Superforecasting got smart people to make predictions for study of their accuracy:

I managed to recruit 284 serious
professionals, card-carrying experts whose livelihoods involved
analyzing political and economic trends and events... Almost all had
postgraduate training; half had PhDs...

In the end:

the average expert was roughly as accurate as a dart-throwing chimpanzee.

Certain groups of pundits - namely
ideological ones and those with frequent TV appearances - actually did
worse than dart-throwing chimpanzees.

Note: This post has been updated with all
elections tracked, up through December 2022. To see the old graphs with
only data going up to 2018, see here.
For an even older page, with only 2016 data, see here.

To come: an analysis looking at more
predictions made by the site over time (say, a week before an election,
or a month) -- as opposed to data used in the above graph, which just
looks at a single prediction for each candidate, taken on the morning of
the election.

As an aside, here were the biggest upsets
we've tracked (which is unchanged since the 2018 update.) The following
candidates were big underdogs in the betting. All won:

Below is ElectionBettingOdds.com's full
track record, sorted by date. Send any questions or comments to Maxim
Lott at maxim.lott@gmail.com or at @maximlott on Twitter. The graphs above are based on these data.

===================

Track Record - Election Betting Odds

By Maxim Lott

Updated Dec 11, 2022

ElectionBettingOdds.com has tracked some 807 different candidate chances across dozens of races between 2016 and December 2022.

How accurate has this site been?

When a candidate favored by bettors loses,
it's easy to fall into the trap of thinking the bettors were "wrong".
But remember that when bettors give someone, for example, a 75% chance
to win -- that also means the bettors give the person a 25% chance to lose.

The following graph shows how closely the betting tracked reality.

The blue line shows what predictions would
look like if they were perfect. The orange line shows how often bettors'
predictions within different ranges (e.g., between 60-80%) actually
came to pass. Where the orange line differs from the blue line, the
predictions were off somewhat:

We can see that, using these prediction
ranges spanning 20 percentage points, ElectionBettingOdds.com was
extremely accurate. For example, out of all candidates who were given
between a 20 and 40% chance of winning at 8am on election day, 27.6% of
them actually won.

We do also see some systematic "underdog"
bias at the edges. For example, for candidates in the 0% to 20% range,
bettors gave an average prediction of a 5.4% chance. Yet, only 1.7% of
those candidates actually ended up winning. The inverse occurs for
highly-favored candidates, who should have been given even higher odds.
This bias has been in part caused by regulatory trading limits on
bettors at PredictIt, which prevents savvy traders on that market from
making much money by correcting such odds.

The graph below holds predictions to a more precise standard -- 10% intervals, with smaller sample sizes.

We can also see that the accuracy of the
odds has been improving over time, as the number of races tracked has
risen and reduced randomness. For comparison, here was the same graph
after our 2018 update, where some 450 candidate chances had been
tracked:

How does all the above compare to other
outlets? For a detailed comparison to the predictions of the best
forecaster (Nate Silver's 538) see here.
That analysis goes all the way from 2016-2022, and does a precise
apples-to-apples comparison between the two prediction methods.
Considering "experts" and models more broadly, several studies suggest
that betting beats listening only to polls. See this one in Public Opinion Quarterly. Also, a study by the author of Superforecasting got smart people to make predictions for study of their accuracy:

I managed to recruit 284 serious
professionals, card-carrying experts whose livelihoods involved
analyzing political and economic trends and events... Almost all had
postgraduate training; half had PhDs...

In the end:

the average expert was roughly as accurate as a dart-throwing chimpanzee.

Certain groups of pundits - namely
ideological ones and those with frequent TV appearances - actually did
worse than dart-throwing chimpanzees.

Note: This post has been updated with all
elections tracked, up through December 2022. To see the old graphs with
only data going up to 2018, see here.
For an even older page, with only 2016 data, see here.

To come: an analysis looking at more
predictions made by the site over time (say, a week before an election,
or a month) -- as opposed to data used in the above graph, which just
looks at a single prediction for each candidate, taken on the morning of
the election.

As an aside, here were the biggest upsets
we've tracked (which is unchanged since the 2018 update.) The following
candidates were big underdogs in the betting. All won:

Below is ElectionBettingOdds.com's full
track record, sorted by date. Send any questions or comments to Maxim
Lott at maxim.lott@gmail.com or at @maximlott on Twitter. The graphs above are based on these data.

===================