How the weather is calculated
Model version 1.0.0. Every number on the map comes from the four public series below and arithmetic you can reproduce from the published data.
The four signals
Each metro is scored on four monthly series. Nothing proprietary, nothing modelled, nothing purchased.
| Signal | Weight | What it is |
|---|---|---|
| Employment | 40% | Total nonfarm payrolls in the metro, monthly |
| Unemployment | 25% | Metro unemployment rate, inverted so lower is hotter |
| Building permits | 20% | New private housing units authorised, three-month mean |
| Labour force | 15% | People working or looking for work in the metro |
Temperature
Each signal is converted to a percentile of that metro’s own trailing 120 months, using only observations that existed at the time being scored. The temperature is the weighted mean of whichever signals are present, renormalised over what is available.
This is the choice that matters most. San Francisco has more of everything than Raleigh, so ranking cities by level would rank them by size. Scoring each place against itself answers the question a reader actually has: is this market unusually busy for itself right now. Fifty means normal here, not average nationally, and a small hot metro and a large warm one are not comparable in dollars.
Unemployment is inverted, once, explicitly: lower unemployment is a hotter economy. Permits are smoothed on a three-month mean first, because one apartment block moves a mid-sized metro’s monthly permit count by forty percent and nobody can act on that.
Momentum, reported separately
Momentum is the change in temperature over 3 months, in points. It is deliberately not folded into the temperature, because the two disagree constantly and the disagreement is the useful part: a market can be hot and cooling, or cold and warming fast, and a single ranking hides exactly the transitions worth knowing about.
Confidence
Three things reduce it. Coverage: how many of the four signals this metro actually has. Recency: how stale the newest observation is, so a slowing feed stops voting at full strength. Agreement: how far the signals disagree with each other, because four signals pointing different ways is not a confident fifty, it is a volatile market.
A market with low confidence is drawn on the map with a dashed edge rather than a different colour, because changing the fill would corrupt the one thing the fill encodes.
Source health
All 100 series are checked on every build and the build fails if any has stopped updating. Currently 100 live.
Boston nearly shipped a 2024 reading. The convenience series FRED publishes for the Boston-Cambridge-Nashua area stopped updating in December 2024 when the geography was retired, and a retired series does not look retired: it still resolves and still returns thousands of rows. Three of Boston’s four signals were dead. The build now refuses to score a market on a feed that has stopped printing, and Boston is reported on the Boston-Cambridge-Newton MSA instead.
What it was tested against, and what it failed
Tested against what happened next, this describes conditions now and does not predict them. It is published as a description, not a forecast.
The test: at 30 historical months, score every metro using only data available at that time, then look forward 12 months and measure what actually happened. 749 metro-month observations across 25 markets.
| Temperature against next year's employment growth | -0.005 | No relationship |
| Hottest fifth, actual growth over the next year | 2.36% | |
| Coldest fifth, actual growth over the next year | 4.54% | Hot markets grew slower |
| Momentum against next year's employment growth | 0.354 | Positive, but see below |
| Momentum against next year's permit growth | -0.142 | The independent test. Fails. |
Two findings, and both are published because both change how the index should be read.
Temperature does not predict. It is very slightly mean-reverting: the hottest fifth of markets went on to grow employment 2.36% over the following year while the coldest fifth grew 4.54%. Being at a high percentile of your own history mostly means you have somewhere to fall back to. Heat is where activity is, not where it is going.
Momentum looked predictive and did not survive the honest test. It correlates at 0.35 with next year’s employment growth, which is encouraging until you notice that employment is forty percent of the index. Momentum predicting employment is substantially momentum predicting itself. Tested against building permits, a series with a different publisher and different drivers, the correlation is -0.142, which is the wrong sign.
So this is published as a description of current conditions and nothing more. There is no outlook, no forecast and no likely-to-warm label anywhere on the site, and there will not be until a version of this table supports one. The gate is machine-read from the dataset on every build rather than being a matter of editorial restraint.
Model versions
Changing a weight changes every historical score, so the version is stamped into every row and the changelog is published. Weights are not retuned silently.
- 1.0.02026-08-25First public model. Percentile-of-own-history temperature, trailing-quarter momentum, coverage/recency/agreement confidence.
Geographies
Metro definitions follow the current OMB delineations, which were revised in 2023. Several older definitions still exist as published series and are deliberately not used. The full list of series behind every score is in the dataset JSON, along with every verification check this build passed.
Source: U.S. Bureau of Labor Statistics (employment, unemployment, labour force) and U.S. Census Bureau (building permits), via Federal Reserve Bank of St. Louis (FRED). Retrieved August 25, 2026. Analysis by DFX Intelligence.