Confident political forecasts are made constantly and revisited rarely. The failure rate is high enough to be worth explaining rather than merely mocking.
Predictions change the thing predicted
A weather forecast does not alter the weather. A forecast that a candidate is certain to win alters donor behaviour, volunteer effort and turnout among both sides.
This feedback loop means the most widely believed forecasts are the ones most likely to undermine themselves, and the effect cannot be measured cleanly after the fact.
Analysts know this and still cannot correct for it, because the size of the response depends on how the forecast is reported rather than on its content.
The sample of comparable events is tiny
National elections happen every few years. Even a century of them yields a small number of observations, and each occurred under different rules, demographics and economic conditions.
Statistical methods designed for large datasets are therefore being applied to a handful of cases. Apparent patterns are fragile, and several well-known ones have reversed within a single decade.
Adding more variables makes this worse rather than better. With few observations and many candidate explanations, a model can fit the past closely while carrying almost no predictive weight.
Incentives reward confidence over calibration
A hedged forecast is accurate more often and gets far less attention. A specific, confident call is memorable, quotable and rewarded regardless of whether it holds.
Because nobody keeps score systematically, the cost of being wrong is low and the benefit of being loud is immediate. The incentive gradient points one way.
Forecasters who publish their track records tend to sound noticeably more cautious than those who do not, which is itself informative.
Continuity is mistaken for causation
Politics contains genuine regularities: incumbency advantages, the effect of economic conditions on approval, and the tendency of a governing party to lose ground at midterm contests.
These regularities are real but weak, and they are routinely treated as laws. When several of them point the same way, stated confidence compounds even though the evidence has not.
Structural change quietly erodes them as well. A relationship that held while party coalitions were stable can weaken considerably once those coalitions begin to rearrange themselves.
How to read a forecast usefully
The useful content of a prediction is its reasoning, not its conclusion. A stated causal chain can be checked against events as they unfold.
Probability language deserves attention too: a stated one-in-four outcome happening is not a failed forecast, and treating it as one discourages honest uncertainty.
Readers who ask what would have to be true for a forecast to be wrong extract more from it than readers who only remember the number.