Methodology
How the data is built
We keep every original file, record where each value came from, and label each figure with its source and as-of date. Here is how that works.
1. Four layers, kept separate
| Layer | Rule |
|---|---|
| Raw source archive | Official downloads, stored as received with the download date. Append-only: never edited, never overwritten by a newer release. |
| Processed versions | Cleaned, dated outputs of each processing run, kept so earlier results can be reproduced. |
| Canonical current data | The current best record for each place, community or builder. It can supersede older values, but it keeps a link back to every raw input. |
| Published extracts | What you see on this site. It can always be rebuilt from the canonical data, and this whole preview is generated by one build script. |
2. Geographic identifiers come first
Records are joined on official codes whenever they exist: 2-digit state FIPS, 5-digit county FIPS, 7-digit Census place GEOIDs, 5-digit ZCTAs and 5-digit CBSA codes. We only match on names when no code is available, and we flag every name-based match. Known pitfalls we handle explicitly:
- ZCTA is not the same as ZIP Code. Census ZCTAs approximate USPS delivery areas but don't match them. When our assigned state for a ZCTA disagrees with the USPS ZIP prefix, we hold the record back rather than guess.
- New England rents are set by town. HUD publishes Fair Market Rents by town in six New England states, so county figures there are our population-weighted averages of HUD's town rows. We label them that way.
- Connecticut changed its county equivalents. Since 2022, Census uses nine planning regions instead of eight counties. Any data still keyed to the old county codes won't join until it's converted, and we show those gaps instead of hiding them.
- Vintages differ. ACS 5-year estimates, Gazetteer years, HUD fiscal years and permit years are different time frames, so every column carries its own label.
3. Calculations we add
When we compute something the source agency doesn't publish, like a statewide Fair Market Rent, we name it as a USA Houses LLC calculation and explain the method next to the number. Today's calculations are:
- State and county 2-bedroom FMR: the mean of HUD's county or town rows, weighted by the population field in HUD's own file.
- Permit ranks and shares: computed from Census Building Permits Survey metro totals.
- Counts: distinct GEOIDs, or distinct records in our catalog.
4. Our own research
The new-construction catalog is compiled by USA Houses LLC from builder and developer websites and, where noted, from builders directly. It isn't government data and it isn't MLS data. We never republish MLS or RPR data. Every record's location precision is recorded, and we're backfilling a dated source on every record. We don't invent community, builder or price details.
5. Checks on every build
A data-check script runs against each build. It verifies source row counts against the source registry, confirms that geographic keys are unique and well-formed, confirms that every page carries source labels and a noindex tag, confirms that private-company data we haven't cleared is absent, and confirms that no page promises a product that doesn't exist yet.
6. Known limits
- ACS figures are estimates with margins of error, and those margins aren't shown yet.
- The state ACS extract doesn't record its fetch date or table IDs. It will be re-fetched and re-verified directly from the Census API before any paid use.
- The Building Permits file doesn't say whether its 2025 annual figures are preliminary or revised.
- HUD's revised FY2026 FMR workbook in our archive can't be opened. The FY2026 figures shown come from HUD's original FY2026 file.
Questions about a figure? Ask us, and we'll show you exactly where it came from.