In 1884 Sergeant John Park Finley of the US Army Signal Service made experimental tornado predictions, twice a day, for 18 districts east of the Rockies. By his own scoring, 96.6 percent of 2,803 predictions came true.

The geologist Grove Karl Gilbert went through the counts the same year. A tornado came just 51 times. So a forecaster who said "no tornado" every single time would have been right 98.2 percent of the time, knowing nothing.

That is the trap with rare events: "it won't happen" is almost always right, and worth nothing. Gilbert scored only the hard part. When Finley said "tornado", one came 28 times in 100; by luck alone it would have been about 2. Finley was good. His percentage was useless.

Next time something is 99 percent accurate, what would "always no" have scored?

Sources: J. P. Finley, "Tornado predictions", American Meteorological Journal 1, 1884; G. K. Gilbert, "Finley's tornado predictions", American Meteorological Journal 1, 1884; Allan Murphy, "The Finley Affair: A Signal Event in the History of Forecast Verification", Weather and Forecasting 11(1), 1996.
Sounds: Freesound, CC0 (felix.blume); our own, made in code (earshot, CC0).