Evaluating prediction intervals
Three panels show the same forecast and the same 20 true values with a nominal 80% interval of different width. Interval A is narrow (width 6.0) and covers 50% of the values; interval B covers 100% with a width of 26.0; interval C covers 80% with a width of 14.0. A good interval is calibrated (coverage close to its target) and sharp (as narrow as possible), as interval C. The Winkler score combines both in a single number, where lower is better: 19.0 for C, against 25.1 for A and 26.0 for B.
Evaluating prediction intervals
1
Three intervals
2
Coverage
3
Width
4
Both
skforecast.org
0.2 s