Seeing cancellations coming, and giving the quiet weeks a reason to book

The situation

Spring and fall are the desert’s good seasons, and the lodges sold those weekends months in advance. Then, every year, a wave of cancellations arrived 10 to 20 days before arrival: plans changed, a group split, a flight connection fell apart. The rooms usually stayed empty. Waitlist calls happened on Monday mornings, by hand, and by the time someone reached the third name, the guest had booked elsewhere.

The other half of the calendar had the opposite problem. Summer midweeks are hot, the weeks after the winter holidays are quiet, and the packages meant to fill them were PDFs that hadn’t changed in three years. The team’s main lever was a discount, and they disliked pulling it.

What we tried first

We started with a rule the team already half-believed: long lead time plus a flexible rate means risk. It flagged nearly half the book. Nobody can call half the book, so by week two the list was ignored.

On the shoulder side, we modeled a broad midweek discount from past data. It mostly moved guests who would have come anyway from Saturday to Tuesday at a lower rate. It added very few nights.

Both failures pointed the same way. The team needed fewer, better flags with reasons attached, and the soft weeks needed a reason to visit, not a lower price.

What we built

A cancellation risk score that explains itself. Signals scores every reservation daily and shows the top reasons beside the number: booked 160 days out on a flexible rate, no deposit yet, a party of six split across two rooms, a channel with a high historical change rate. The score is a prompt for a person, not a decision.

Three bands, three playbooks. Low risk: nothing. Medium: a warm reconfirmation 21 days out with practical arrival notes. High: the same message plus a call task for the team. Released rooms go to the waitlist in order through Booking Flow, with a 24-hour hold on each offer.

Packages built around reasons to come. For each forecast soft midweek, Signals drafts two or three package ideas from what the desert offers that week: darker skies near the new moon, cool sunrise hours, quiet trails. The reservations manager picks, edits or rejects. Booking Flow then requests holds from local guides and restaurants, parses their confirmations and issues vouchers.

Guardrails on data and outreach. The model never sees names or free-text notes. Outreach goes only to guests who opted into pre-arrival messages, and every template was written by the team, in its own voice.

What changed

Monday mornings got shorter. Instead of a long list, the team worked a short one with reasons, and knew what to say: a guest with a split party got a question about room arrangements, not a generic reminder.

Resale of late-cancelled room-nights more than doubled, mostly because waitlist offers went out within an hour of a release instead of the following Monday. Package components confirmed without anyone chasing them about seven times in ten.

The midweek occupancy gain was real but modest, and a year-over-year comparison carries market noise. We report it with that caveat attached.

What we’d do differently

Model group blocks separately. Wedding and retreat blocks cancel differently and skewed the first model. We split them out in week six; they deserved their own model from day one.

Onboard package suppliers earlier. One night-sky guide confirms only by phone. We built a call-back task rather than forcing automation, which was right, but we learned it in week eight instead of week two.

Show the misses, not just the hits. Our first backtest review showed the cancellations the model caught. Trust grew faster once we walked the team through the ones it missed.