Occupancy Prediction

Overview

Short-term rental hosts & platforms need accurate annual revenue projections, but raw scraped calendar data only distinguishes “available” from “blocked” days — it doesn’t say why a day is blocked. This project classifies every calendar day into one of three occupancy states — booked (a guest reservation), offline (host-blocked for maintenance or personal use), or vacant (open & available) — so revenue estimates reflect true earning potential rather than raw availability.

The Challenge

• Vacant days expose a nightly price; booked & offline days look identical in the raw data — both simply appear “blocked”
• No direct signal for why a day is blocked: a real guest reservation vs. the host blocking dates for maintenance or personal use
• Severe class imbalance — roughly 65% vacant, 30% booked, 5% offline
• Time-series constrained: predictions can only use features known up to that point in time, with no leakage from the future
• Must generalize to brand-new properties never seen during training, not just properties already in the dataset
• Very limited first-party ground truth to validate predictions against

Approach

• Hierarchical classifier combining property-level & market-level time-series features
• Train/test split by property (not by date) to evaluate performance on unseen listings
• Benchmarked against — & designed to emulate — an existing but imperfect third-party data source

Results

• Achieved 4% MAPE on annual revenue projections derived from the classified calendars