How often is our interval right?
Every market gets its own prediction interval calibrated against its own held-out sales at each retrain. Most markets target 85%; the higher-volume ones target 90%. Because calibration actively tunes the band to its target, healthy markets should sit within a fraction of a point of goal — read this page as “did calibration run, on how much data, and how tight are the bands”, one row per market. A market drifting more than a couple of points from target is the anomaly worth noticing.
Coverage = the fraction of held-out recent sales that closed inside our prediction interval, measured on the calibration split at each market's last retrain — not a live day-by-day tracker of individual listings. Conformal calibration tunes each band toward its target on this split, so values very close to target are the expected healthy state (that's the method working, not a coincidence); the number to watch is the signed gap in the right column and the sample size. Forward coverage — how the bands hold up on sales that closed after publication — is a separate, harder test we're instrumenting now and will publish on this page.
Each row is a fleet-aggregate number for the whole market — it can differ by price segment. In practice, budget-tier coverage tends to run below the market average shown here, while premium-tier intervals are sometimes materially tighter than nominal. Treat the row above as a market-level summary, not a guarantee for any one price band.