How we read the evidence

Recorded sales, estimates and asking prices

Official completed transactions describe historical sales. Contributor asking prices are separate marketplace observations. A model estimate is published only after time-based testing and promotion checks; it is never called an HDB or bank valuation.

Comparable sales

HDB and non-landed private resale are separate cohorts. We start with the same location, HDB flat type or compatible private class, single-unit resale and floor area within 15%. Specified tenure, remaining lease and floor bands constrain the cohort. We start at 12 months and expand to 24 if fewer than five observations remain. We do not silently substitute nearby locations.

What the range means

At least five compatible sales are required for a descriptive median and interquartile range. The lower and upper quartiles describe the middle half of observed prices; they are not a confidence or prediction interval. Small samples are shown as insufficient evidence. Unknown unit condition, facing and renovations remain limitations.

Model validation

Model candidates use official transaction snapshots, separate train/calibration/test periods, and held-out error, coverage and interval-width diagnostics. Historical validation is labelled retrospective unless source publication-time history is known. Failing or undersupported models stay unpublished. Submitted asking prices and sponsorship never train or influence the model.

Freshness and corrections

Complete validated source snapshots are promoted atomically. Corrections and withdrawals replace current evidence. A failed import preserves the last successful snapshot and its real age. Raw medians can change when the mix of sold properties changes.

Policy version 5 September 2026. See source rights and attribution and corrections.