Skip to main content
cohost
Cohost Analytics · accuracy

Our forecast error, published.

Every STR data vendor claims accuracy. None of them show their backtests. Here are ours: per market, our revenue forecast against Inside Airbnb's per-listing revenue estimates across the whole revenue distribution — including the markets where we're not good enough yet.

Market Median error Worst single miss Reality inside our range Sample State
Chicago 26.6% 82.9% 69% 16 listings Validated · <30%
Nashville, TN 27.7% 147.5% 50% 16 listings Validated · <30%
Denver 29.1% 74.9% 38% 16 listings Validated · <30%
Broward County, FL 34.0% 79.8% 38% 16 listings Wide band — improving
New Orleans, LA 39.7% 125.3% 44% 16 listings Wide band — improving
San Diego, CA 43.1% 101.5% 38% 16 listings Wide band — improving
Asheville 43.9% 84.8% 50% 16 listings Wide band — improving
Austin 60.9% 95.9% 31% 16 listings Wide band — improving

Backtest: stratified sample across each market's revenue quartiles (not just top performers), forecast = median of nearby comparable listings, "realized" = Inside Airbnb's review-evidenced estimated_revenue_l365d per listing. "Reality inside our range" is how often the actual value landed within the p25–p75 band we showed — a calibrated band should be near 50%.

Why publish this

Comp-based STR forecasting carries ±20–30% error in liquid markets — that's the industry's open secret, and per-listing error is larger than the market-average numbers vendors quote. In November 2025 the category leader's revenue projections silently collapsed for paying customers when a comp-selection change went wrong. The fix for an industry built on unaudited point estimates is boring: show the backtest, every market, every model change.

The honest part
  • "Realized" here is Inside Airbnb's review-evidenced estimate, not owner bank statements. When owner-reported revenue lands, it replaces this baseline — same ledger, better source.
  • Wide-band markets stay on the board. Hiding them would defeat the purpose.
  • Numbers regenerate with every model change — they can get worse, and you'd see it here first.