A take. A guess. Some data.
This is Josh’s housing-market toy. It is designed to make the national data easier to talk about, not to promise what your home will be worth.
The weekly opinion
A fixed set of rules reads the daily snapshot. A mortgage-rate pressure score of 70 or more produces “Financing pressure is squeezing buyers.” Otherwise a GH score of at least 55 produces a cautious headline; lower readings produce a mixed headline. Missing inputs or stale core observations replace the stance with an incomplete-data message. Those cutoffs are choices, not proven economic laws.
The first successful daily run in each Monday–Sunday week saves an issue. Later daily runs do not rewrite it. Each issue stores its evidence, observation dates, score, wording, and method version. A missed week stays missed; no retrospective weekly issues are fabricated. The initial rule set is outlook-1.0.
What the GH score means
Higher means more pressure according to our chosen directions. Each numerical signal is transformed, then placed in its available trailing 3,650-day distribution. Ties get half weight. Downward-pressure signals reverse the percentile. At least 20 observations are required. Some series have much less than ten years of usable history.
| Layer | Weight |
|---|---|
| Headline keyword sentiment | 15% |
| Financing and housing equities | 25% |
| Housing behavior | 30% |
| Economic foundation | 20% |
| Reported housing outcomes | 10% |
Signals within each numerical bucket are averaged equally. Financing is 60% and equities 40% of their combined layer. Missing whole layers cause the remaining weights to be rescaled. GH averages the current composite and up to two previous recorded composites. The pressure bands are below 40, 40–54.9, 55–69.9, and 70+. Neither those thresholds nor the weights have been fitted to future housing outcomes.
The 26 numerical inputs sit alongside a headline keyword heuristic. They are not 26 independent votes; some overlap economically. “Leading” and “lagging” are organizational labels, not proof of a measured lead time. The index has recorded observations since July 2026, so it does not have an established long-term predictive record.
Putting a number on a guess
Target: the mean monthly US new single-family home sales rate in the named quarter or calendar year. Source: Census / HUD via FRED, HSN1F. Units are thousands of homes at a seasonally adjusted annual rate. The UI multiplies by 1,000. A 670,000 annualized quarterly average does not mean 670,000 homes sold that quarter.
Two candidate rules: keep the latest monthly rate unchanged, or use the mean of the latest three months. For each horizon, we select the lower mean absolute error on retrospective windows beginning in 2000 and ending before 2020. We separately report performance on origins from 2020 onward. Every simulated prediction uses only observations at or before its origin, but the observations themselves are revised values downloaded today. We have not reconstructed historical data vintages or publication lags.
The rough range is the point estimate plus or minus the 80th percentile of absolute training errors, floored at zero. It is not an assured probability interval. Held-out coverage and average error are reported alongside a no-change baseline. Monthly evaluation windows overlap; they are not independent experiments. This is not a backtest of GH and does not establish investment usefulness.
The record
The next-quarter guess is issued once per target quarter. The next-calendar-year guess is issued once per target year. Both remain fixed. Daily runs count available outcome months; when all target months are present, they preserve that first observed complete result. Absolute error and comparison with the original no-change baseline are displayed. Later revisions do not overwrite the saved result.
Original forecast JSON, full issue-time input snapshots, and result JSON live under data/outlook/. Files are created exclusively, never overwritten by the publisher. The aggregate index is rebuilt for display. This is a transparent file archive, not cryptographic proof against edits by the site owner. Corrections must be appended and explained, never silently substituted.
Sources and limits
The evidence cards link directly to their source series, with observation dates. Observation month is not publication date. Realtor.com inventory and days-on-market cover listings; new-home sales cover new single-family homes. These populations differ. The national data cannot determine a local asking price, an individual mortgage quote, or a household’s best financial decision.
Player guesses are stored locally in the browser. They are a private toy, not a public contest or audited leaderboard. Clearing storage clears the guess.
Version outlook-1.0 · September 29, 2026. Download current record.