Work out click prediction value instantly with clear inputs, formula shown and shareable results.
In ad and feed ranking, the click model's predicted probability is multiplied by value to set bids and to forecast revenue, so calibration matters as much as ranking quality. A model with a perfect ranking but a predicted CTR 14 percent above reality overstates revenue by 14 percent and overbids by the same margin, which is why the predicted-over-actual ratio is monitored continuously.
Revenue from CTR
revenue = impressions x CTR x revenue per click; calibration ratio = predicted CTR / actual CTR
Post-hoc calibration: Platt scaling or isotonic regression fitted on recent data maps predicted scores onto observed frequencies and is cheap to refresh daily.
Training on a subsample of non-clicks shifts the base rate, so predictions must be corrected back with the known sampling ratio before they are used as probabilities.