July 15, 2026
Forecasts, feelings, and chaos
The well-calibrated Bayesian [pdf]
Math nerds lose it over an old paper saying good predictors should match reality
TLDR: This paper says a good probability forecaster’s numbers should match what really happens over time, and that this matters deeply for how we think about rational belief. Commenters split between calling it a timeless classic, joking that it’s ancient history, and turning the title into a music meme.
A dusty 1982 statistics paper just got the internet doing what it does best: turning a very serious idea into a comment-section variety show. The paper’s core claim is surprisingly relatable even if the title sounds like it escaped from a graduate seminar: if someone keeps saying there’s a 30% chance of something, then over time that thing should happen about 30% of the time. In other words, a good probability guesser shouldn’t just sound smart — their guesses should line up with real life. The spicy part is that the author says a fully consistent Bayesian thinker would actually expect to be this well-calibrated, with potentially “destructive implications” for bigger theories about rational belief.
But the real fireworks came from the crowd. One camp treated the paper like a celebrity reappearance: “This is a classic!” declared one commenter, while another dryly dropped “(1982)” like a tiny timestamp grenade, reminding everyone this “news” is older than many readers. Then came the jokes. One person wondered if the title was a wink to The Well-Tempered Clavier, turning a probability paper into accidental sheet music discourse. Another heroic explainer broke the concept down with a rain forecast example, then swerved into a beautifully weird joke about being “calibrated” while still being useless. And, because no comment thread is complete without a dramatic historical quote, Cromwell’s Rule got summoned with: “think it possible that you may be mistaken.” Honestly? That may be the whole vibe of the thread.
Key Points
- •The article examines subjective probability forecasting using weather prediction as its main example.
- •It defines calibration as agreement between forecast probabilities and long-run observed frequencies for events assigned those probabilities.
- •The paper treats forecasting within a coherent Bayesian framework in which forecasts summarize conditional distributions given current information.
- •Dawid states that calibration applies to both event probabilities and credible interval forecasts.
- •The article’s central theorem is that a coherent Bayesian expects to be well calibrated, which Dawid says has destructive implications for the theory of coherence.