Airbnb’s Dynamic Pricing System and Machine‑Learning Platform (Aerosolve)

Airbnb needed a better way to help hosts set rental prices, so in 2012 it began developing an automated pricing tool that evolved into a dynamic pricing system released in June 2022, which provides daily price suggestions based on constantly changing market conditions.

The system combines a general pricing algorithm with a unique machine‑learning component that learns from both historical data and human intuition, allowing it to adapt to unexpected market features.

Unlike simpler pricing services such as eBay, Airbnb’s challenge is complex because each of its millions of listings is unique, with varying attributes, host requirements, seasonal trends, and local events that affect demand.

Initially, in 2013, Airbnb launched a static pricing tool that used a few key features (e.g., number of rooms, beds, amenities) but could not adjust for time‑sensitive factors or location nuances.

Since mid‑2021, the company has rebuilt the tool to interact with users, learn from errors, and provide demand‑driven price adjustments, eventually open‑sourcing the underlying machine‑learning platform called Aerosolve.

Three illustrative scenarios are presented: a World Cup‑driven surge in Brazil, a unique Scottish castle with historic data, and a Paris two‑bedroom apartment facing high demand, each highlighting the need for nuanced pricing inputs beyond basic size and location.

The overall architecture extracts key property attributes when a host adds a listing, finds comparable successful listings in the same area, and generates a centered price suggestion, using similarity, recency, and location data.

Similarity data includes quantifiable attributes, capacity, property type (apartment, castle, yurt), and review count, which surprisingly has a strong impact on price.

Recency considers seasonal demand and local events; for example, prices in Austin rise during SXSW and ACL festivals.

Location handling evolved from simple radius‑based similarity to detailed neighborhood boundaries drawn by cartographers, allowing fine‑grained price differentiation even within a few hundred meters (e.g., Greenwich vs. London Docklands).

Dynamic pricing now updates nightly suggestions for each host based on the planned rental dates, similar to airline and hotel revenue‑management systems, and leverages years of historical data despite the computational cost.

The machine‑learning model is a classification model that predicts whether a listing will be booked at a given price, using hundreds of features such as breakfast inclusion, private bathroom, and days‑to‑booking, and continuously retrains based on observed outcomes.

Feature importance is automatically adjusted; for instance, photo style (bright living rooms vs. warm bedrooms) influences booking likelihood, and new features like "days before check‑in" are added as they prove predictive.

The system also refines neighborhood maps to capture "micro‑neighborhoods" where price behavior differs sharply from traditional boundaries, as illustrated by the London micro‑neighborhood map.

Today, the tool provides price suggestions for Airbnb listings worldwide, and the underlying Aerosolve platform has been open‑sourced to enable engineers outside the travel industry to build similar machine‑learning applications.

Author: Dan Hill, former Airbnb product lead for pricing algorithms, now focusing on technology and product development.

Source: 大数据文摘