AI data market
Hospitality operational data investment: room moves over guest files
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Hospitality operational data investment should target coordination records, not guest profiles. In a public set of 119,390 bookings from two Portuguese hotels, 14,115 of 75,166 completed stays (18.8%) ended in a different room type than reserved, and 74.7% of those show no booking amendment. Room codes are anonymized and no field says why, so the move is visible and the decision is not. That gap, not guest volume, sets value.
Table overview
Hospitality operational data
Key takeaways
- In the public Hotel Booking Demand data, 18.8% of 75,166 completed stays ended in a different room type than reserved: 25.4% at the resort hotel and 14.7% at the city hotel.
- Three in four of those room moves (74.7%) show no booking amendment, and anonymized room codes hide whether a guest moved up or down.
- The decision behind a move, its reason, owner and guest response, lives in notes and messages, which is where a supplier's preparation burden sits.
- Guest identities are a liability: the FTC's 2024 Marriott order followed breaches affecting more than 344 million customers and requires limits on how long personal data is kept.
- Ask which entity controls each property's systems, owner, manager or brand, before asking how many rooms a portfolio has.
What would a hospitality data investor actually be buying?
The investable part of hospitality operations data is the record of coordination: how a hotel reassigned rooms, routed a guest's problem to housekeeping or maintenance, recovered a stay that went wrong and settled the folio. Guest identities are not the asset. They are the part a licensing deal works hardest to remove.
Coordination is the job itself. O*NET's profile of hotel, motel and resort desk clerks lists tasks such as contacting housekeeping or maintenance when guests report problems, recording guest comments or complaints and referring guests to managers, and telling housekeeping when rooms have been vacated and are ready for cleaning. Each is a handoff between people and systems with a result someone can check.
To judge the opportunity, an investor needs the same evidence for each task family: a buyer task, a structured trace, a verifiable outcome, a rights holder able to say yes, and some measure of utility. Public data supplies parts of the first three, as the bookings recount below shows, and almost none of the last.
One stay in five changed room type: what 119,390 public bookings show
Room moves are common enough to matter. In the Hotel Booking Demand datasets that Nuno Antonio, Ana de Almeida and Luis Nunes published in Data in Brief in 2019, covering arrivals from July 2015 to August 2017 at a resort hotel in the Algarve and a city hotel in Lisbon, 14,115 of 75,166 completed stays (18.8%) were assigned a different room type from the one reserved.
The rate depends on the property. The resort hotel moved 25.4% of its completed stays and the city hotel 14.7%, so a portfolio's mix of property types changes how many such cases it holds before anything else does. Of all moved stays, 10,547 (74.7%) show zero booking amendments, which suggests the move was made at room assignment or check-in rather than through a logged amendment to the reservation. Another 802 bookings that were canceled or never arrived also carry a different assigned room.
The counts below come from the authors' two files as redistributed by the TidyTuesday project, recounted for this article:
| Measure | Resort hotel | City hotel | Both hotels |
|---|---|---|---|
| Bookings | 40,060 | 79,330 | 119,390 |
| Completed stays (status Check-Out) | 28,938 | 46,228 | 75,166 |
| Stays assigned a different room type | 7,341 (25.4%) | 6,774 (14.7%) | 14,115 (18.8%) |
| Moved stays with zero booking amendments | 5,530 (75.3%) | 5,017 (74.1%) | 10,547 (74.7%) |
| Moved stays with at least one special request | 3,216 (43.8%) | 3,014 (44.5%) | 6,230 (44.1%) |
| Bookings with any days on the waiting list | 255 (0.6%) | 3,443 (4.3%) | 3,698 (3.1%) |
Why a room-move record cannot tell a rescue from a slip
Because the data records the move and hides its meaning. The data dictionary says the assigned room type sometimes differs from the reserved one for hotel operation reasons, such as overbooking, or at the customer's request, and it shows room types as letter codes for anonymity. No field says which reason applied, who decided or what the guest was told.
Three details sharpen the gap. The letter codes cannot be ranked, so an upgrade and a downgrade look the same. Two codes, I and K, appear only as assigned rooms and never as reserved ones, and the data cannot say what they are. The booking-changes field counts amendments without recording who made them or why.
That is the general shape of hospitality operations data. The event lands in the property management system; the decision lives in shift notes, messages, maintenance tickets and the folio. A supplier's real work, and an investor's real diligence question, is whether those pieces can be joined case by case.
Guest information is a liability to minimize, not inventory to sell
Guest data belongs on the cost side of the ledger. The Federal Trade Commission's order against Marriott and Starwood, finalized in December 2024, followed three breaches that the agency said affected more than 344 million customers worldwide and exposed passport information, payment card numbers and loyalty numbers. The order requires an information security program, limits on how long personal data is kept and a way for U.S. customers to ask for deletion.
NIST's Generative AI Profile (NIST AI 600-1) names data privacy among the risks particular to generative AI, describing impacts from leakage and unauthorized use, disclosure or de-anonymization of personal or sensitive information. A dataset that carries guest names, contacts, passports or cards into model training passes that risk to every party downstream.
The public bookings data shows that the alternative works. Its authors removed the elements that identified the hotel or the customer, and they present the files as useful for revenue management and machine learning research; the recount above needed nothing more. A coordination dataset can be scoped the same way: events, timings, roles and outcomes, with guest identity removed at the source rather than offered as a feature.
Who can license a hotel's operating records?
Often not the company that owns the building. Hotels split ownership, management and brand, and the systems that hold operating records may be run by the manager or the franchisor rather than the owner.
Sunstone Hotel Investors, a hotel owner, told investors in its annual report for 2019 that it and its third-party managers and franchisors rely on information technology networks and systems to process and store customer information, including large volumes of guests' personally identifiable information. An owner in that position may hold neither the records nor the authority to license them.
So the first diligence question for a hospitality portfolio is not the room count. It is which entity controls the property management system, the work-order system and the guest messaging tool at each property, what the management and franchise agreements say about operating data, and who signs.
Hospitality task scorecard: five families and their missing evidence
Room-status handoffs and room moves score highest at 7 of 8; folio disputes score lowest at 3, because they carry the most payment and identity data. Each family gets 0 to 2 points on four criteria: public evidence of the task, a structured trace, a checkable outcome and guest-data exposure, where 2 means least exposed. The scores are editorial readings of the sources in this article, and no row has a public measure of model utility.
Families, traces and gaps:
| Task family | Coordination a buyer wants | Trace a system keeps | Missing evidence | Score of 8 (task, trace, outcome, exposure) |
|---|---|---|---|---|
| Room-status handoffs | Front desk, housekeeping and maintenance agree a room is vacated, cleaned and ready | Status changes with timestamps | Who released the room and whether the next guest reported a problem | 7: 2, 2, 1, 2 |
| Room moves and relocations | Assign another room when the booked one is unavailable or unsuitable | Reserved and assigned room type; 18.8% of public completed stays differ | Reason, direction of the move and the guest's response | 7: 2, 2, 1, 2 |
| Service recovery after a complaint | Record the complaint, route it, choose a remedy within authority, confirm the fix | Complaint notes and any credit on the folio | Remedy approval, fix confirmation and a repeat-contact check | 5: 2, 1, 1, 1 |
| Group and event changes | Change a room block, catering and deposits together | Block, contract and deposit records | Task not documented in the sources here; who approved each change | 5: 0, 2, 2, 1 |
| Folio adjustments and billing disputes | Decide a charge correction or refund and post it | Folio lines and adjustments | Reason and approver, plus card and identity data that must be removed | 3: 0, 2, 1, 0 |
Do more rooms and more bookings mean more licensable supply?
No. In the public data, extra volume mostly adds routine and canceled bookings: 43,017 of 119,390 bookings (36.0%) were canceled, and 49,376 of the 61,051 completed stays that kept their room type (80.9%) show no amendment at all. The decision-rich cases are a minority, and their share depends on the type of property.
Three things shrink supply further. The records that grow fastest with guests, such as profiles, messages and payment details, carry the most personal data. Control of the systems may sit with managers or brands rather than owners. And the reasons behind moves and remedies may exist only in free-text notes that take time to join and redact. Supply is best counted as cases with a joined decision and outcome, property by property, with the counting rule stated.
What could make the hospitality data thesis wrong?
Four risks belong in any memo on this sector:
- Privacy enforcement could narrow scope. The FTC's Marriott order already requires limits on retention; similar orders or state rules could leave fewer historical records to license.
- Brands and managers could keep the records. Where franchisors or management companies run the systems, as Sunstone's filing describes, independent owners have less to offer and supply concentrates with a few operators.
- Role-played data could be good enough. MultiWOZ, a public corpus of about 10,000 written human-human dialogues across domains that include hotels, was collected with crowd workers playing visitors from task descriptions; if data like that trains capable booking agents, real records add less.
- Public evidence is thin. The bookings recount covers two Portuguese hotels from 2015 to 2017; room-move rates elsewhere, and whether richer records improve a model, are unmeasured.
Diligence questions for a hospitality records supplier, SourceX included
Ask questions whose answers are counts and documents, and ask SourceX the same ones.
SourceX licenses selected business records to AI developers on a company's behalf and carries the work from qualification and rights review through contracting, preparation, delivery and payment; the company approves scope and terms and signs before records are delivered to SourceX. It is raising from accredited investors under Rule 506(c), and it shares the company and commercial metrics it does not publish with investors who request materials through its investors page.
Six questions for any hospitality records supplier:
- Which entity controls each system that holds the records, and who signs for licensing at each property?
- How many room moves, recoveries or status handoffs fall in the offered period, and how many have a reason, an owner and an outcome joined to them?
- How are guest identity, contact, passport and payment fields removed, and at which step?
- What share of decision context is free text, and how is it redacted before delivery?
- How do exception rates differ between property types in the portfolio?
- Has any buyer tested the records on a held-out task, and what did its acceptance test require?
Methods note: the bookings recount and the scorecard
Counts come from H1.csv and H2.csv in the rfordatascience/tidytuesday repository, folder data/2020/2020-02-11, at commit 6994076 (cloned October 27, 2026; SHA-256 prefixes 6f38fda1 and 5107090f), recounted with hotel_room_change_count.py. Rules as applied: every row is a booking; text fields are trimmed of spaces; a room-type change is a row whose assigned room type differs from its reserved room type; a completed stay has reservation status Check-Out; no amendment means booking changes equal zero; a waiting-list booking has days in waiting list above zero. No sampling or randomness was used, and the row totals match the 40,060 and 79,330 observations the authors report.
Scorecard points are editorial, 0 to 2 per criterion, read from the cited sources on October 27, 2026. The Data in Brief, O*NET, FTC, NIST and Sunstone points were verified through searches limited to each publisher's site, so a reviewer should open them before release. Not measured: room-move rates at other hotels, the share of moves with a recoverable reason, preparation cost, prices, buyer demand and model utility. No SourceX data was used.
Frequently asked questions
Is hotel operations data useful for AI training?
Its coordination records can be: room moves, status handoffs and service recoveries with their reasons and outcomes. Public data shows such events are frequent, with 18.8% of completed stays in one dataset assigned a different room type, but no public study measures their training value.
Should a hotel license guest data?
Guest identity should be removed, not licensed. The useful signal sits in events, timings, roles and outcomes, and the FTC's Marriott order shows the cost of holding personal data longer than needed. Counsel should review any dataset that touches personal information.
Why does property type matter to hospitality data supply?
Exception rates differ by property. In the public data the resort hotel moved 25.4% of completed stays to a different room type and the city hotel 14.7%, so two portfolios with the same room count can hold very different numbers of decision cases.
Who signs a data license for a branded or managed hotel?
It depends on who controls the systems and what the management and franchise agreements say. Owners may rely on managers' and franchisors' systems, so confirm the controlling entity and signing authority property by property before scoping a dataset.
What evidence should an investor request from a hospitality data supplier?
Counts of joined decision cases per property with the counting rule, the method for removing guest identity, and one sample case traced from event to outcome. Then ask whether any buyer has tested the records against a held-out task.
Sources
- Antonio, Almeida and Nunes (Data in Brief, 2019) describe two hotel booking datasets, a resort hotel (40,060 observations) and a city hotel (79,330) in Portugal, arrivals July 2015 to August 2017, with hotel and customer identifiers removed. Source
- The TidyTuesday redistribution of the Hotel Booking Demand files (H1.csv, H2.csv) and data dictionary, used for this article's recount: 14,115 of 75,166 completed stays assigned a different room type. Source
- O*NET's profile of hotel, motel and resort desk clerks lists coordination tasks such as contacting housekeeping or maintenance when guests report problems and recording guest complaints. Source
- The FTC finalized an order with Marriott and Starwood in December 2024 over breaches affecting more than 344 million customers, requiring a security program, data retention limits and deletion requests. Source
- NIST AI 600-1 lists data privacy among generative AI risks, covering leakage and unauthorized use, disclosure or de-anonymization of personal information. Source
- Sunstone Hotel Investors' 10-K for 2019 says it and its third-party managers and franchisors rely on IT systems that store customer information, including large volumes of guests' personal information. Source
- MultiWOZ is a public collection of about 10,000 written human-human dialogues across multiple domains, with hotel among its benchmark domains, collected with crowd workers working from task descriptions. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.