A demand-signal model that beats a fitted AR rent-growth forecast.
HomeStar forecasts metro rent growth from labor, job-postings, and housing-flow data, and beats a fitted autoregressive (AR) baseline out-of-sample across the 108 largest US metros. It produces a better rent-growth forecast — an input to underwriting under a known rate environment — not a buy/sell signal.
The improvement is largest at 3 months (+17.2%), consistent across market sizes, and comes mostly from the housing-flow inputs (which add +9.5% at 3 months). All figures are out-of-sample, measured against a baseline built from each metro’s own recent rent history.
For each metro and month, the model predicts rent growth 3, 6, and 12 months out from its own recent history (the AR baseline) plus demand signals: employment and JOLTS quits/hires (labor), the job-postings index, and Zillow housing-flow (for-sale inventory, new listings, days-to-pending). Labor on its own does not beat AR; postings and housing-flow are what add forecast skill, most of it at the short horizon.
Pre-registered and out-of-sample on two axes at once: an expanding-window walk-forward through time, and a held-out-metro split (the model is scored on metros it never trained on; the two axes agree). Skill is the reduction in forecast error (RMSE) versus the baseline; each gain is tested to confirm it's statistically real (not chance), adjusted for testing several horizons at once, and reported with confidence intervals. The harness and its output are committed and reproducible.
- Scope: the 108 largest US metros (full labor + postings + housing-flow coverage); smaller metros are out of scope.
- Window: ~5 years (2021–2026), one major regime shift — cross-regime durability is untested.
- Pooled: skill is reported across metros, not per metro.
- Rent (target): Zillow Observed Rent Index (ZORI).
- Labor: BLS metro employment and state JOLTS quits/hires (via FRED/BLS).
- Job postings: Indeed Hiring Lab postings index (via FRED).
- Housing flow: Zillow Research for-sale inventory, new listings, days-to-pending (downloaded research CSVs).
- Live concession layer: a de-biased concession read, collected via Bright Data's web-data platform — see Data Sources.