DwellSignal
Methodology

How we count

The definitions behind every DwellSignal number, what measured and modeled mean on a page, and where a figure is withheld rather than guessed.

Every figuremeasured, modeled or withheld, never guessed

Every figure on a DwellSignal page traces back to a specific listing, permit, or open dataset, and every figure carries one of three labels. This page says what the labels mean and what has to be true before a number is shown.

Three labels

Measured. Observed directly. A nightly rate is the price the listing shows. A permit match is a listing sitting at an address that holds a current registration. Booked nights are nights we watched flip from available to unavailable between two of our own calendar snapshots.

Modeled. Produced by a model we have tested out of sample and whose error we publish. Occupancy for a place without measured listings comes from a small-area model trained on places that have them. A modeled figure is served as a band, never a point, and the band is calibrated to hold the true value at the stated rate.

Not measured. Nothing is shown. The page says why: too few listings, no permit registry for the city, no forward calendar readings, or a figure that did not reconcile with the others.

What has to be true before a number appears

Denominators, named

Occupancy is served on the underwriting basis: booked nights over all nights in the window. The vendor-comparable basis, booked over available nights with host blocks removed, runs about ten points higher on the same bookings, and where we show it we say so. We do not drop listings with no recent booking from an average; a dead listing is information about the market.

Sources

Listings, calendars, and reviews come from our own market dataset and from Inside Airbnb; permits from city and county registries and published ordinances, matched at the address level; rents and home values from Zillow research data; hazards from FEMA's National Risk Index; demographics from the Census. Each page names the dataset behind each figure and the date it was taken.

Errors we publish

The nightly-rate model's leave-one-city-out error is 16.9%. The occupancy model's is 21.8%, and it is served only where its calibration holds. When a definition changes, the change is dated in the methodology changelog and the affected pages say "updated".

Every number here is produced by the same rules the product uses and is withheld where the sample cannot carry it. The definitions live in the methodology.
Newsletter

Movers, rule changes, and the numbers behind them

One email a week. Withheld figures included.