There odd actually something in this. Nate Silver’s book ‘The Signal and The Noise’ has a whole chapter on the phenomena of weather forecasting and introduced the interesting idea (at least to me) of systematic bias. I have asked Claude to summarise this: apologies if this reads like AI slop but I think it does relate to this discussion and I haven’t got an hour to remember and summarise it myself.
Here is what Claude said:
In The Signal and the Noise, Nate Silver examines a striking and counterintuitive finding about the accuracy of weather forecasts: they are systematically biased, and the direction of that bias depends on who is doing the forecasting and for whom.
The core finding
Commercial weather forecasters — particularly those serving the general public, such as television meteorologists — tend to exhibit a wet bias: they overpredict the probability of rain relative to what the actual frequency of rainfall warrants. If a forecaster says there is a 70% chance of rain, it should rain on roughly 70% of such occasions. In practice, it rains considerably less often than commercial forecasters predict.
Why the wet bias exists
Silver's explanation is essentially one of asymmetric loss functions — a concept borrowed from decision theory. The consequences of a false negative (failing to warn of rain that arrives) are experienced as considerably worse, from the audience's perspective, than a false positive (predicting rain that does not materialise). A person who is unexpectedly caught in a downpour is far more aggrieved than one who carries an umbrella unnecessarily. Forecasters, being attuned to their audiences' reactions, respond rationally to this asymmetry by shading their predictions toward rain. There is also a reputational element: a forecaster blamed for ruining a picnic suffers more conspicuously than one who merely caused mild over-preparation.
The contrasting case: the National Weather Service
Silver notes that the National Weather Service (NWS), the US government's official forecasting body, is considerably better calibrated — closer to true probabilistic accuracy. Silver attributes this partly to the fact that the NWS is insulated from direct commercial pressure and audience sentiment. Its forecasters are evaluated against objective accuracy benchmarks rather than public approval.
The broader lesson
Silver uses this example to illustrate a theme running throughout the book: that incentives shape forecasts, often invisibly. A forecast is never purely a technical output; it is produced by a human or institution embedded in a set of relationships, pressures, and incentives that systematically pull predictions in particular directions. Understanding a forecast well requires understanding who made it, and why they might prefer to err in one direction rather than another. It is, in essence, a lesson in epistemic humility about what forecasts are actually measuring — and a warning against taking probabilistic statements at face value without interrogating the institutional context from which they emerge.