AI Smart Pricing for Short-Term Rentals

Nightly Price Suggestions for Rental Hosts

Design the system that suggests a price for every open night on every rental listing, learns from which nights booked, and applies the suggestions inside limits the host sets.

Overview

A rental host has to set a price for every night on the calendar, often a year ahead. Too high and the night stays empty; too low and it books at once and the host earns less than they could have. A smart-pricing service suggests a price for each open night and, if the host opts in, applies it automatically inside a minimum and maximum the host chooses.

Airbnb described its own version in a KDD 2018 paper (Ye and colleagues). Hosts stay in control: Price Tips shows a calendar with the predicted chance of a booking for each night, and Smart Pricing lets a host set a price range so that new suggestions inside it are applied automatically. The paper splits the work into three parts: a model that predicts whether a night will book, a strategy model that turns that into a price, and a final layer of adjustments for hosting goals and special events. This design follows the same split.

Questions to ask first

Who acts on the price? — The host, either by accepting a suggestion or by turning on automatic pricing with a min and max
How far ahead? — Every open night in the next 365 days
How fresh must prices be? — A full recompute nightly; nights in the next 30 days again every 4 hours
What signals exist? — Listing details, the calendar, bookings, search and view counts, and a feed of local events
What is the goal? — More host revenue over the year, not the fullest calendar
Who owns the calendar? — The listing service; this system only proposes prices and asks it to write them

What you will be able to do

Split pricing into a booking-probability model, a price strategy and a guardrail layer, and say what each one owns.
Explain why maximizing price times booking probability is fragile, and what Airbnb reported when it tried it.
Train a price model from booked and unbooked nights with an asymmetric loss, and measure it with price decrease recall and booking regret.
Price a new listing by borrowing from its market, and fix gap nights that no minimum stay can fill.
Push prices to the calendar so that a host's own edits always win.

Key concepts

Booking probability
Demand curve
Expected revenue
Asymmetric loss
Price decrease recall
Booking regret
Shrinkage to market priors
Lead time
Gap nights
Host guardrails

Functional requirements

Suggest a nightly price for every open night in the next 365 days of each listing
Show the host the predicted chance of a booking at the current price and at the suggested price
When automatic pricing is on, apply suggestions only within the host's min and max, and never over a night the host priced by hand
Refresh near-term nights when bookings, search activity or a new local event change demand
Flag gap nights that the current minimum stay makes unbookable, and suggest a fix
Record every suggestion with its model version, so outcomes can be joined back for training

Non-functional requirements

Nightly run finishes within a 4-hour window (our target)
Nights in the next 30 days refreshed at least every 4 hours (our target)
Host dashboard reads under 300 ms at p99 (our target); suggestions are precomputed
A failed run leaves yesterday's prices in place; nothing is written half-done
Every automatic price change is auditable: who or what set it, when, and from which model

Interview questions

Why not simply pick the price that maximizes price × booking probability?
How do you train a pricing model when you never see the optimal price?
How do you price a listing that has no history?
How do local events get into the price?
How do you make sure you never overwrite a price the host set?
What happens if the nightly run fails halfway?
How would you test a new pricing strategy?