Observed Probability Calibration Review

How to use Observed Probability Calibration Review

Compare each probability with its observed zero/one outcome. Equal-width bins retain counts, mean predictions and outcome rates; exact weighted bin gaps yield ECE. All positive/negative pairs supply tied half-credit AUC, while log loss remains an approximate numerical score.

1. Enter model json (schema and units below).
2. Choose the stated convention and check the units and supported model.
3. Calculate, inspect the result, then copy or download it.

Good to know

• Plain decimal JSON numbers or strings, at most six fractional places, magnitude at most 1,000,000,000; raw numeric spellings and duplicate object members are checked. Exact rational arithmetic is capped at 8,192 bits. Text is at most 30,000 characters; a local JSON download preserves exact fractions or full stored approximate numbers. 1–1,000 observations and 1–50 bins. Outcomes are 0/1 and probabilities lie in [0,1]. A wrong zero-probability outcome makes meanLogLoss null with an unbounded-penalty count. One-class AUC is null. No fitting, clinical risk advice, deployment recommendation or data-leakage guarantee.
• Local calculation only; no input persistence or external lookup. Text/JSON is limited to 30000 characters and output to 1000000 characters. Numeric model inputs and precision/work limits are stated on this page. Approximate JSON numbers retain full stored values; exact arithmetic outputs retain decimal or fraction strings. Unsupported states fail explicitly.