What is Calibration?
How closely stated confidence matches the real hit rate of calls made at that confidence.
A model is calibrated when the calls it makes at 70% are right about 70% of the time. The calibration chart plots stated confidence against observed frequency, with the number of calls behind each bin.
This is the chart to read before trusting a confidence number. It is built from graded calls only, so it is sparse at the start and fills in as the track record grows.
How our model uses it
Each forecast carries the observed hit rate of past calls at its confidence bin, once that bin has graded calls behind it.
Related terms
- ConfidenceOur model
The probability of the call's direction, as a number the site publishes next to every forecast.
- Hit rateOur model
The share of graded calls that were correct, by horizon and by model version.
- Track recordOur model
The public log of every graded call, including the losses, with the calibration behind the confidence.