# Reading Probabilistic Forecasts A forecast saying a candidate has a 30% chance of winning is not a prediction that they will lose. It is a claim that in a long run of comparable situations, that candidate wins about three times in ten. Thirty percent events happen constantly. They are roughly as likely as a random NFL field goal attempt from 55 yards. The public — and much of the press — reads these numbers as pronouncements rather than as distributions, and the resulting misinterpretation has consequences that outlast any single election. ## The three standard misreadings **Treating anything above 50% as a prediction.** A 70/30 forecast becomes "X will win" in a headline, and when the 30% outcome occurs the forecast is declared wrong. But a forecaster who says 70% and is right exactly 70% of the time is *perfectly calibrated*. The only way to evaluate probabilistic forecasts is across many of them, using a proper scoring rule such as the Brier score. **A single election cannot falsify a probabilistic forecast**, which is a genuinely uncomfortable property and not a dodge. **Confusing vote-share margin with win probability.** These are different quantities and they move differently. A candidate leading by 2 points with high uncertainty may have a lower win probability than one leading by 1 point with low uncertainty. Late in a race, win probabilities move much faster than vote margins, because the same margin becomes more decisive as the time remaining for it to change shrinks. Reporting that treats a jump from 65% to 80% as evidence of a surge is usually reporting the clock, not the polls. **Ignoring correlated state errors.** Model outputs differ enormously depending on whether state errors are treated as independent. Independent errors make a coordinated multi-state miss nearly impossible and produce very confident forecasts. Correlated errors — the empirically correct treatment, since polling misses share causes across states (see [[Poll Aggregation]]) — produce much wider distributions. **Most of the difference between competing forecasts in a close election is this parameter, not the polls.** ## Why the misreadings are politically load-bearing **Complacency.** A widely reported high win probability plausibly suppresses turnout and donations among supporters of the favorite. The forecast becomes an input to the event it is forecasting — a feedback loop no purely statistical treatment can escape. **The betrayal narrative.** When a low-probability outcome occurs, "the models were wrong" becomes "the models were rigged." This is the on-ramp from a technical misunderstanding to [[Election Denialism]]. The forecasters' defense — that they said it could happen, and gave it a specific non-trivial probability — is mathematically correct and rhetorically useless, because the audience never held the number as a distribution in the first place. **Selective invocation.** Forecasts are cited as authoritative when favorable and dismissed as elite modeling when not, often by the same speaker within a cycle. This is [[Manufacturing Doubt]] applied to a probabilistic object: because a probabilistic forecast is never strictly refuted by a single outcome, it is also never strictly vindicated, which leaves permanent room for motivated reinterpretation. ## What a well-communicated forecast looks like Lead with the distribution rather than the headline probability. Show the range of plausible outcomes and how wide it is. State the correlated-error assumption explicitly. Report the model's historical calibration. Present the probability alongside a familiar comparison of the same magnitude, because "30%" and "about the odds of rain tomorrow when the forecast says 30%" land very differently in the same reader. Most importantly: **make the uncertainty the headline, not the footnote.** The single number is what gets quoted, so if the single number is the point estimate, the uncertainty will not survive transmission. ## Related [[Poll Aggregation]] · [[Herding in Polling]] · [[Likely-Voter Models]] · [[Risk vs Uncertainty]] · [[Election Denialism]] · [[Mathematics and Power Hub]]