Forecasts and their track record

How far ahead the data we already hold can predict — deaths, disease, demand — and what the record of those forecasts shows.

Projection has the same character with a longer horizon. Official projections of life expectancy have been wrong in the same direction for decades — too low, until the 2010s, when in several countries they became too high as improvement stalled — and the question is whether anyone forecasts mortality better than the offices, and how one would know. The models are well described: extrapolative, cause-specific, expert-judgement, ensembles, and now machine-learning approaches that claim more from the same data. Their track records are less well described, because forecasts are seldom scored against outturns in public and the horizon that matters is decades. What the evidence does show is that the errors come mostly from turning points — the stall, the pandemic — that no extrapolative model anticipates, and that cause-specific and expert-judgement models fared no better at them. The same pattern holds wherever a projection is made from existing data rather than from new observation. Disease incidence, care demand and the cost of both are projected on the same kinds of model, with the same scarcity of published scoring against what followed. A projection of one disease's future burden belongs on that disease's page; what belongs here is the question of whether such projections have held up, and which kinds have held up better.

Status

building

Direction

mixed

Would change our view

A published comparison scoring competing forecasts against outturns over a decade or more; a method that demonstrably anticipated a turning point rather than fitting it afterwards; a national office publishing probabilistic projections as its main output.

Gaps

Public forecast archives in most countries; any scoring of the actuarial and insurance models, which are private; forecasts above age 90, where the data are thinnest and the stakes highest.