The Winkler score (or interval score) grades a forecast range — like a 90% confidence band — against the price that actually happened. It rewards bands that are both narrow and honest, and lower is better.
For a band with lower edge L, upper edge U and a coverage level (say 90%, so alpha = 0.10), the score is the width (U − L) plus a miss penalty of (2/alpha) times how far outside the band the price landed. Inside the band, only the width counts. Example (illustrative): a band of $95–$105 (width 10) at a 90% level. If the stock closes at $100, the score is just 10 — covered and tight. If it closes at $110 (5 above the top), the penalty is (2 / 0.10) × 5 = 100, for a total of 110. One badly missed forecast dominates many good ones — it punishes overconfident bands hard.
A simple "did the price land in the band?" check can be gamed by making every band absurdly wide — a forecast of "somewhere between $0 and $1,000" is never wrong and never useful. The Winkler score charges for width, so a model cannot buy coverage with vagueness. It is the honest single-number answer to "were these bands both calibrated and sharp?"
The Winkler score is a standard public statistic, not a proprietary edge. We show it in the ticker "Model Quality" bar so you can judge our forecast bands the way a statistician would. How they actually score is reported against a committed backtest on our calibration page, alongside CI90 coverage and the Brier score. None of this is investment advice.
Lower is better — it is a distance-like penalty, not a percentage, so it only has meaning relative to a benchmark or the same model over time.
Coverage only asks whether the price landed inside the band. The Winkler score also charges for how wide the band was, so a model can't score well just by making every band enormous.
No — it's a public statistic from the forecasting literature (Winkler, 1972). We use it because it is a standard, checkable way to grade prediction intervals.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.