Like many others, it took me a while to digest the results of the 2024 U.S. presidential election. But there's one more important thing to do: to analyze how the predictions I made actually fared. And I finally got around to doing that. the predictions I made I generally made three different predictions over the course of ~1.5 months: ① based on weighted high-quality polls only (I downloaded those from 538); ② adjusted by the full average polling bias from the 2016 and 2020 elections; and ③ adjusted by just half the 2016/20 polling bias. half What was my reasoning for integrating the polling bias from 2016 and 2020? "I'm not considering elections before that because ever since Trump entered the stage, election dynamics have significantly changed. Old rules don't apply anymore. 2024 will be much more similar to 2020 & 2016 than to any election before that." (See my very first update from September 22, 2024.) update from September 22, 2024 update from September 22, 2024 And why did I include the one with the half polling bias? In my update from October 9, 2024, I wrote: "It's most probably just as unrealistic to assume there'll be negligible systematic polling error in the swing states this year (①) as to assume it'll be as big as in 2016/20 (②). The truth will most probably lie somewhere in the middle." update from October 9, 2024 update from October 9, 2024 Beginning of November -- based on very reasonable-sounding arguments -- I also started looking into whether the polling errors from 2012 and 2022 might be relevant to consider. Unfortunately, those arguments turned out to be completely, utterly wrong, so we'll exclude predictions ④ and ⑤ in this post-mortem analysis. What predictions ① thru ③ did was to predict the winning margin in the seven swing states, which then yielded a probability of winning in each state for the candidates, expected electoral votes (EVs), and, combined (and based on a Monte Carlo simulation), an overall chance of winning the whole thing. What I will focus on here is to look at how far off my margin predictions were from the actual margins in the swing states, because those were the ones on which all other components of my forecast were based. I have chosen to do so with three common approaches: directional accuracy, mean absolute error (MAE), and root mean square error (RMSE). While MAE simply calculates the absolute difference between predicted margins and actual margins and then averages these values, RMSE penalizes large errors more heavily. I will also do MAE and RMSE for an equal weighting of the seven swing states as well as for an EV-based weighting each. Equal weighting treats each state prediction with equal importance, focusing purely on predictive accuracy regardless of strategic importance. EV-based weighting, however, better reflects the electoral significance of each prediction. For both, MAE and RMSE, lower is better. All of that being said, let's get to it. Let's see how my predictions fared. Metric

Prediction ①

Prediction ②

Prediction ③



Directional Accuracy

42.9%

100% ✓

100% ✓



MAE (equal weights)

2.37

2.11

1.01 ✓



RMSE (equal weights)

2.46

2.32

1.32 ✓



MAE (EV-weighted)

2.26

2.11

0.81 ✓



RMSE (EV-weighted)

2.35

2.32

1.06 ✓ Metric

Prediction ①

Prediction ②

Prediction ③



Directional Accuracy

42.9%

100% ✓

100% ✓



MAE (equal weights)

2.37

2.11

1.01 ✓



RMSE (equal weights)

2.46

2.32

1.32 ✓



MAE (EV-weighted)

2.26

2.11

0.81 ✓



RMSE (EV-weighted)

2.35

2.32

1.06 ✓ Metric

Prediction ①

Prediction ②

Prediction ③ Metric Metric Prediction ① Prediction ① Prediction ② Prediction ② Prediction ③ Prediction ③ Directional Accuracy

42.9%

100% ✓

100% ✓ Directional Accuracy Directional Accuracy 42.9% 42.9% 100% ✓ 100% ✓ 100% ✓ 100% ✓ MAE (equal weights)

2.37

2.11

1.01 ✓ MAE (equal weights) MAE (equal weights) 2.37 2.37 2.11 2.11 1.01 ✓ 1.01 ✓ RMSE (equal weights)

2.46

2.32

1.32 ✓ RMSE (equal weights) RMSE (equal weights) 2.46 2.46 2.32 2.32 1.32 ✓ 1.32 ✓ MAE (EV-weighted)

2.26

2.11

0.81 ✓ MAE (EV-weighted) MAE (EV-weighted) 2.26 2.26 2.11 2.11 0.81 ✓ 0.81 ✓ RMSE (EV-weighted)

2.35

2.32

1.06 ✓ RMSE (EV-weighted) RMSE (EV-weighted) 2.35 2.35 2.32 2.32 1.06 ✓ 1.06 ✓ As we can see, no matter which method we use and how we weigh the individual swing states, prediciton ③ easily outperforms the other two (except for directional accuracy, the most basic measure, that looks at how many states were predicted correctly independent of margin). Since all three predictions are based on the same polls with the same weightings and only differ in how much they were adjusted for which polling biases, this shows that the assumptions I originally made were reasonable: Elections with Trump as a candidate are unlike other elections; pollsters are still bad a capturing support for Trump (due to the "shy Trump voter" etc.); pollsters still improved a bit, though; simple models (take the average polling error from 2016/20 and just cut it in half) mostly do the trick. To conclude this entire topic (until we start anew less than 4 years from now), one more piece of further reading: There are reports that a French bettor on Polymarket engaged so-called neighbor polling (very roughly speaking, asking "who do you think your neighbor will vote for?" instead of "who will you vote for?"), thereby predicting everything correctly. Prof. Andrew Gelman has written a very interesting piece on this. everything very interesting piece on this very interesting piece on this Alright then, see you in 3.5 years. 👋🏻

No Testing Is (Sometimes) Better Than Some Testing

I Tried to Predict the 2024 Election—Here’s What I Got Right (and Totally Wrong)

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Untitled Story

A Smidge of Behavioral Science at the End of the Year

What Are Convolution Neural Networks? [ELI5]

The Noonification: Have U Been Pwned? (1/12/2023)

Goldman Sachs, Data Lineage, and Harry Potter Spells

People are still crazy about Python after twenty-five years

A Smidge of Behavioral Science at the End of the Year

What Are Convolution Neural Networks? [ELI5]

The Noonification: Have U Been Pwned? (1/12/2023)

Goldman Sachs, Data Lineage, and Harry Potter Spells

People are still crazy about Python after twenty-five years

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps