Robert Winkler

Who they are

This page records the public professional record only. Nothing adverse is recorded and no motive is attributed.

His affiliation and wider career were not verified for this page, so none is asserted here. The record used is the 1977 paper, on which he is the second author. It sits in the literature on subjective probability forecasts, which asks whether people who state probabilities state them honestly, and it is an example of that literature being applied to a quantity that resolves every day.

Key papers

Murphy, A. H., & Winkler, R. L. (1977), 'Reliability of Subjective Probability Forecasts of Precipitation and Temperature', Journal of the Royal Statistical Society, Series C (Applied Statistics), 26(1), 41-47, doi:10.2307/2346866. Read at the level of the published abstract.

His wider publication record was not surveyed for this page.

Key ideas

The paper summarises findings from operational and experimental programmes in the United States. The forecasts examined were subjective probability forecasts of whether precipitation would occur and of daily maximum and minimum temperatures. The conclusion was that weather forecasters can formulate such forecasts in a reliable manner.

Reliability here has a precise meaning, set out under forecast calibration: the stated probabilities should match the frequencies with which events occur. A forecaster who says thirty per cent should be right about three times in ten.

The reason the weather is the standard case is practical. It resolves every day, so a forecaster gets a steady supply of outcomes to be scored against. A retailer using a forecast built on that record faces a separate question, which is how many units a given temperature moves. That question is the subject of What a Weather Forecast Does for Sales.

Reading the result well means keeping its scope in view. It concerns one period, one country and two kinds of quantity, and it was read here at the level of its abstract.

How to read the result

The paper is the work of two authors, and this page is about the second. The record verified for this site is the paper alone, so the useful thing to say about Robert Winkler's place in it is what its question was. The question was whether people who state probabilities state them in a way that matches the world, and the weather was an unusually good place to ask it because it resolves every day.

A reader using the result should separate three claims. One is descriptive: in the American programmes summarised, forecasters gave probabilities for rain and temperature, and those probabilities were reliable. A second is conditional: if a retailer trusts such a forecast, it has a dependable input. The third is not in the paper at all: that the input will improve a sales forecast. The third belongs to the retail studies, such as the 2020 work of Florian Badorf and the 2017 work of Sebastian Steinker, which found gains in forecast accuracy of different sizes and over different horizons.

Keeping the claims apart also guards against a common slip, which is to treat a forecaster's confidence as evidence about demand. A reliable thirty per cent chance of rain tells a planner how often the day will be wet. It does not tell the planner how many coats that wet day will sell.

What the record does not show

The record does not describe his later work on forecasting or decision analysis, and none is asserted here. Nor does it say how the paper's authors divided the work. The abstract, which is what was read, names both authors and one conclusion, and this page keeps to that. Anyone wanting more should start from the paper itself, linked on its study page, and treat the paper's date as a boundary: it dates from 1977 and says nothing of today's ensemble methods, which What a Weather Forecast Does for Sales draws from a separate review.

---