A demand model produces a point forecast of 1,200 units for next week. Using split conformal prediction with a 90% target coverage on a calibration set, the interval is [980, 1,420].
What does the 90% guarantee mean?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Conformal prediction takes the model's residuals on a held-out calibration set, finds the 90th-percentile absolute error, and adds it around each new prediction. The resulting intervals cover the truth 90% of the time in the long run, whatever the model (GBDT, neural net, anything) and whatever the error shape — as long as new data is exchangeable with the calibration data. That is a promise a planner can put a safety stock behind.
The guarantee is about the interval containing the outcome, not about the point forecast being right.
Conformal does not aim for narrow intervals; it aims for honest ones. Width is whatever the calibration residuals dictate.
R² describes fit; coverage describes how often intervals contain the truth. They are unrelated numbers.