The Signal and the Noise

Silver examines forecasting across domains where records exist, and argues that the difference between good and bad forecasting is a specific set of practices rather than better models or more data.

The domains he covers include weather forecasting, which he identifies as one of the genuine success stories with measurable improvement over decades; earthquakes, where he argues prediction is not currently possible and the attempts have produced false confidence; economics, where he documents that forecasters do not publish confidence intervals and that the intervals they do have are far wider than users assume; baseball, poker, chess, political polling, epidemiology and climate.

His central prescription is Bayesian reasoning, meaning starting from a prior probability, updating it in proportion to how diagnostic new evidence is, and stating beliefs as probabilities rather than as positions. He argues that the failure to state a prior is what makes most forecasting unfalsifiable and unimprovable. His analysis of the 2008 crisis identifies four specific failure points: the rating agencies' models assumed regional rather than national default correlation, the housing boom was treated as sustainable, the effects of the collapse were underestimated, and the models were built on data from a period containing no comparable event.

The book was published in 2012 and Silver's own subsequent forecasting record has been mixed, particularly around 2016, which he addressed by noting his model gave a higher probability than others. The critique that Bayesian methods depend on the prior, which is chosen rather than derived, applies to his prescription. The domain comparisons are the most valuable part.