How well does any of this work?
Every number on this site comes from a method that has been tested by forecasting the past without looking at it. This page gathers every one of those tests, the good and the bad. The short version: strong El Niño precipitation along the southern tier has real, modest skill; most other things are close to zero, and a few are worse than doing nothing. From April 1, 2027, the page also grades this winter's forecasts against what actually happened.
Every skill number, on one axis
Leave-one-out, as publishedA skill score says how much better a forecast did than a simple reference, in percent of the way to perfect: 0 means no better than the reference, below 0 means worse. RPSS scores the three odds (below, near, above normal) against always saying one in three. Brier skill scores yes/no lines against the 1991–2020 base rate. Bars are 95% intervals from resampling whole winters, where they exist. Every forecast here was made for a winter without that winter's data.
The Season Book backtest
Brier skill against the 1991–2020 base rate, by family
Reliability: forecast vs what happened
Book probabilities in ten bins against the share of lines that came true. On the diagonal is perfect.