Where Can I Get A Hotel Booking Dataset For Machine Learning Data: Dataset Structure and Field Coverage

One record answers this question outright: Hotel Booking Demand (Kaggle – 119k Portuguese Hotel Bookings), cataloged by Datadory at quality score 9. It ships as a single CSV of 119,390 rows by 32 columns — roughly 17 MB — one row per booking at daily arrival-date resolution, drawn from exactly two properties: H1, the Resort Hotel, and H2, the City Hotel, both in Portugal, with arrivals spanning July 2015 through August 2017.

The schema walks the whole booking lifecycle, which is why feature engineering rarely stalls for lack of inputs:

1. Timing — lead_time (days between PMS entry and arrival), arrival_date_year, arrival_date_month, arrival_date_week_number, arrival_date_day_of_month, stays_in_weekend_nights and stays_in_week_nights.
2. Party and product — adults, children, babies, the meal package (Undefined/SC, BB bed & breakfast, HB half board, FB full board), reserved_room_type versus assigned_room_type, and required_car_parking_spaces.
3. Commercial — market_segment and distribution_channel (TA = Travel Agents, TO = Tour Operators), agent and company IDs, deposit_type (No Deposit / Non Refund / Refundable), adr, and customer_type (Group, Transient, Transient-Party or Contract).
4. Behavior and outcome — is_repeated_guest, previous_cancellations, previous_bookings_not_canceled, booking_changes, days_in_waiting_list, total_of_special_requests, the ISO 3155-3:2013 country of origin, and the reservation_status pair (Canceled, Check-Out, No-Show).

Two structural quirks shape preprocessing. Granularity is one row per booking, not per room-night, so a five-night stay contributes a single blended adr and nightly variance is understated until you expand it. And the two-property sample means country-level generalization is a case study, not a benchmark.

A defensible pipeline chains the microdata to live aggregates in four steps:

1. Train the canceller and rate models on the 119,390-row CSV — classification on is_canceled, regression on adr, using the lead-time, deposit and segment fields above.
2. Attach demand context by joining Eurostat tour_occ_arn2 arrivals and nights onto your geography keys; use the resident/non-resident split to isolate inbound pressure, and mind the documented time-series break between the 2011 and 2012 reference periods.
3. Add supply-side features from Eurostat tour_cap_nat — establishments, bedrooms and bed-places by NACE I551/I552/I553 back to 1990, 24,765 observations, DOI 10.2908/TOUR_CAP_NAT — to normalize demand against room stock.
4. Refresh the picture past the 2017 cutoff with NTTO I-94 monthly releases (through July 2026 preliminary) and OpenStreetMap Overpass queries, which reflect the live database rather than a snapshot.

Three constraints recur in every serious write-up of this corpus, and all three come straight from its structure:

Within those bounds the corpus remains the most-cited public dataset for hospitality demand modelling, and pairing it with the official series above produces features that survive review.

Start with the hotels resorts cruise lines data guide, the pillar mapping all 12 cataloged datasets in this industry and how they fit together. Sibling clusters go deeper on the two modeling angles closest to this one: hotel booking cancellation dataset works the labeling problem end to end, and ADR revenue management dataset treats the pricing side.

For the records themselves, the dataset pages document every field and access path: Hotel Booking Demand (Kaggle – 119k Portuguese Hotel Bookings), Eurostat – Nights Spent at Tourist Accommodation (tour_occ_arn2) and OpenStreetMap Hotels via Overpass Turbo. Browse the full slice from the hotels-resorts-cruise-lines data hub, see the ranked shortlists on best hotels-resorts-cruise-lines datasets and free hotels-resorts-cruise-lines datasets, or check how data scientists apply these sources on the hotels, resorts & cruise lines data for data scientists page.