Wanderhome

Thousands of vacation homes are listed. Some make money and some don’t, and the industry’s explanation for which is which was mostly folklore. Wanderhome was built to replace the folklore with a regression.

What the model did

We bought the performance data for every short-term rental in the United States and joined it to MLS listing and sold-home data. Homes were bucketed into 40 segments by sleeping capacity, bedrooms and bathrooms, then analysed at segment × ZIP code — with automated clustering of similarly-behaving nearby ZIPs where a single one had too few comparables to be meaningful.

It produced three things:

What each characteristic is worth. Not “a hot tub is nice” but the measured effect of each attribute on average daily rate and on occupancy, separately — because they are not the same and they frequently point in opposite directions. A sloped driveway in a snow market suppresses rate. A jacuzzi lifts occupancy more than it lifts rate.

What the operational settings should be. Does Superhost status actually pay? Is a five-star rating necessary or does 4.7 perform identically? How many photos is optimal? How strict should the cancellation policy be? These were treated as parameters to be estimated per segment and ZIP, not as best practices to be copied.

What to buy, and what to spend on it after. A prevalence model produced three tiers of attributes: the minimum set required to hit baseline performance in that segment and ZIP, the attributes correlated with high performers that don’t actually drive rate, and the attributes that do. That third tier was the renovation budget.

The honest constraint

The most useful thing the model told us was where its own ceiling was. Nightly rate is bounded by the neighbourhood — even the best house in a ZIP won’t book at twice the going rate. We could reliably run 10–20% above comparable homes in the same segment and ZIP, and pushing past that traded rate for occupancy at a loss. Knowing the size of the edge is what makes it investable.

Proof

We bought one home in Big Bear Lake, California and ran it for a year unmodified, then for a year with the model’s recommended upgrades.

Same house, same market, same year-over-year demand environment. Revenue doubled and a losing asset became a profitable one. We reached Superhost within six months of launching the brand.

Alongside the model sat the operating stack: MySQL and Python for the pipeline on AWS, a React Native guest app with smart-home device control, noise sensors reporting to an operations centre, and an iPad field tool that let the acquisition team score a candidate house on site against the model.

What this is evidence of: that we build models that make a specific, expensive decision — which asset to buy and what to spend on it — and then go put our own money behind the output.