We merged 35,870 Los Angeles Airbnb listings with Zillow neighborhood home values, then compared four modeling approaches to predict each listing's estimated annual revenue.
A random sample of listings across the LA basin, colored by estimated annual revenue. Hover a dot for details.
Number of listings by estimated annual revenue, in $10K bins
Almost half of all listings earn under $10K a year, while the mean ($23K) is pulled up by a long tail. That's why we trimmed the 1st and 99th percentiles before modeling.
Median estimated annual revenue by room type
Entire homes make up 77% of listings and earn about 2.6× more than private rooms.
Median estimated annual revenue, neighborhoods with at least 75 listings
Avalon (Catalina Island) earns about 3× more than any other neighborhood. Beach cities and the Santa Monica Mountains fill out the rest of the top five.
Each dot is a neighborhood with a Zillow match and ≥30 listings: median home value vs. median estimated annual revenue
Neighborhood home value correlates only weakly with revenue (r = 0.14 across 120 neighborhoods). Listing-level factors like price, availability and room type carry far more signal.
R² on the held-out test set (higher is better)
Test-set error in dollars of annual revenue
| Model | MAE | RMSE | R² |
|---|
The neural net trained on all features roughly halved MAE compared to the tree ensembles, to about $10.8K per listing.
Neural net test R² by listing price tier
The model is strong for mid- and high-priced listings (R² 0.75–0.83) but fails on the top tier. Luxury listings are rare and volatile, which points to a separate model for that segment as the next step.
Estimated annual revenue = nightly price × (365 − available nights), from Inside Airbnb LA listings.
Parsed prices, dropped invalid rows, removed the 1st/99th percentile tails and minimum stays over 365 nights.
Exact, partial, city-level, then fuzzy name matching linked 76% of listings to a neighborhood home value. The rest were median-imputed.
Scaled numeric + one-hot categorical features. Compared linear regression, random forest, gradient boosting and tuned neural nets on MAE, RMSE and R².