Skip to content

Systems and records

Hotel multi-property policy dataset: keeping each property's rules in training data

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Hotel training data preserves property-specific rules when every booking, request and decision is tied to the property, the rate plan, the policy version in force when the guest booked and the stay dates, not to the brand's current policy page. Build a coverage matrix of properties, policy types and versions, link each case to its cells, and record where law or an exception overrides the house rule.

System export

Hotel multi property policy

This page covers hotel multi-property policy dataset: keeping each property's rules in training data.Illustrative

Key takeaways

  • Resolve every rule from four keys: property, rate plan, policy version on the booking date and stay dates.
  • Apply a fixed precedence: reservation terms, then rate plan, then property version, then brand default.
  • Anchor purchase-time rules to the booking date and operational rules to the stay date.
  • Record legal overrides, such as service animal and accessible room duties, as their own column.
  • Publish coverage counts per property, policy type and version, and grade models on the reservation end state.

How should hotel training data preserve property-specific rules within one group?

Tie every booking, guest request and staff decision to four keys: the property, the rate plan, the policy version in force when the guest booked and the stay dates. The rule that governs a case is resolved from those keys, not from the brand's policy page or from whichever property the dataset happens to have most of. If a case cannot be tied to all four, mark the rule as unresolved instead of filling it in.

Hotel groups are a hard case for training data because the name on the building says little about the rules inside it. Properties under one brand can differ in check-in age, pet acceptance, deposits, parking, early departure charges and cancellation deadlines, and the same property sells rate plans with different refund terms on the same night. Ownership and management add layers: an owner, a management company and a brand can each set or approve parts of the rulebook, and each may keep its own records.

A model trained on pooled examples learns an average hotel that does not exist. It will tell a guest at one property the pet rule from another, or apply a flexible rate's cancellation window to a non-refundable booking. The fix is structural: a coverage matrix that says which properties, policy types and versions the data actually covers, and a case record that points to exactly one cell of it.

Why the brand policy page is the wrong source for labels

The brand page states defaults; the reservation carries the terms the guest actually bought. Labels taken from the brand page are wrong whenever a property or rate plan differs, which is often the case that matters most.

Distribution standards already treat these rules as rate-level and reservation-level facts. Google's Hotel Prices transaction message reference expresses refundability with a Refundable element that can be set on a result, on package data or on individual rates, with the rate level overriding the package level and the package level overriding the result level. A full refund window is given as refundable_until_days before check-in plus an optional refundable_until_time in the hotel's local time, which defaults to midnight when omitted.

General vocabularies do the same for stay times. In the schema.org vocabulary, checkinTime is the earliest someone may check into a lodging establishment and checkoutTime the latest someone may check out, and both properties apply to a LodgingBusiness and to a LodgingReservation. A property default and a reservation's own times can therefore legitimately disagree, and the dataset should store both.

The practical consequence is a rule of precedence the labelers apply in order: the specific reservation's terms, then the rate plan, then the property's policy version on the booking date, then the brand default. Record which level supplied each rule in the case, so a reviewer can see why a label was chosen.

Which policy version applied: the booking date or the stay date?

Purchase-time rules are fixed when the guest books; operational rules apply on the day they are used. Mixing the two anchors is the most common way a hotel dataset ends up with labels that look plausible and are wrong.

Cancellation deadlines, deposit and prepayment terms, early departure charges and the total price shown at booking belong to the purchase. A property that tightened its cancellation window in spring should not have that newer rule applied to a winter booking made before the change. Pool hours, parking arrangements, amenity availability and housekeeping schedules belong to the stay date, because guests encounter them as they happen and properties change them for operational reasons.

Price disclosure is a clear example of why the booking-date anchor matters. The FTC's Rule on Unfair or Deceptive Fees, which covers short-term lodging and took effect on May 12, 2025, requires the total price, including mandatory fees such as a resort fee, to be shown. Only three kinds of charges may sit outside it: government charges such as taxes, shipping charges, and fees for optional add-ons, and those must be disclosed before the consumer is asked to pay. Guest disputes about fees on bookings made before and after that date can call for different handling, so each case needs the booking timestamp and the price display record from that moment.

Where the law overrides a property rule

Some property rules give way to legal duties regardless of what the property's policy says, and the dataset must show the override, not just the house rule. Otherwise a model learns to enforce a pet fee on a service animal because the pet policy said so.

The Department of Justice's service animal guidance under the ADA says a hotel cannot charge a deposit, surcharge or cleaning fee for a service animal even when it charges these for pets, though it may charge for damage if it normally charges other guests for that damage. Guests with service animals must be given the same opportunity to reserve any available room and cannot be limited to pet-friendly rooms.

Reservations carry their own duties. The ADA Title III regulations, at 28 CFR 36.302(e), apply to a public accommodation that owns, leases or operates a place of lodging, for reservations made by any means including third parties. They require identifying and describing accessible features in enough detail for a guest to judge whether a room meets their needs, holding accessible rooms until other rooms of that type are rented, and blocking a reserved accessible room so it is removed from all reservation systems.

In the matrix this becomes an override column. Each policy type lists the legal checks that can supersede it, and each case records whether an override applied, so a model learns the order of authority rather than the house rule alone. This page describes the data structure; counsel decides how a specific law applies to a specific property.

The property and policy version coverage matrix

The matrix has one row per policy type and records, for each, what makes it vary, which date anchors its version, the evidence a case must link and the override checks. Fill one matrix for the group, then count cases per property, policy type and version cell to see where the data is thin.

  • Cell: one property, one policy type, one version. A case links to every cell its decision depends on.
  • Coverage count: cases per cell, with empty and thin cells listed rather than hidden.
  • Precedence: reservation terms, then rate plan, then property version, then brand default, recorded per rule.
The property and policy version coverage matrix
Policy typeVaries byVersion anchorEvidence a case must linkOverride or authority check
Cancellation and refund windowProperty, rate plan, channelBooking timestampRate plan code, refund terms text, deadline in property local timeDisruption waiver or manager exception with approver role
Deposit and prepaymentProperty, rate plan, length of stayBooking timestampPayment schedule, deposit taken, refund recordService animal deposit not allowed
Total price and mandatory feesProperty, date rangeBooking timestampPrice shown at booking, fee list, disclosure screen recordFTC fees rule for bookings on or after May 12, 2025
Check-in and check-out timesProperty, room type, loyalty tierStay date, unless set on the reservationProperty default, reservation-level time, late checkout grantStaff authority limit for late checkout
Minimum check-in age and IDProperty, local lawStay datePolicy text version, front desk noteLocal law check by reviewer
PetsProperty, room typeBooking timestamp for fees, stay date for rulesPet policy version, fee charged, room assignedService animals are not pets; no pet fees or room limits
Accessible roomsProperty, room typeBooking timestampAccessible feature description, block status, room assignedReservation duties under 28 CFR 36.302(e)
Parking and amenitiesProperty, seasonStay dateAmenity schedule version, closure noticeDisclosure of closures to booked guests
Early departure and no-showProperty, rate planBooking timestampCharge applied, guest notice, waiver recordDisruption or manager exception with approver role

Worked scenario (illustrative): one traveler, two properties and a canceled flight

This scenario is illustrative. The properties, rates and times are invented to show how the matrix decides a case; they do not describe any real hotel group.

Starting state. A traveler holds two linked reservations in the same group: three nights at property A on a flexible rate, then two nights at property B on a prepaid non-refundable rate. A flight cancellation delays arrival by one day. The traveler messages the group's contact center at 6 pm property A local time, two days before check-in, asking to push both stays back one night.

Inputs from the matrix. Property A's flexible rate allows a full refund until 4 pm local time two days before check-in, under the version in force on the booking date. That deadline passed two hours earlier. Property B's rate is non-refundable, and its early departure rule charges for unused nights. Property A's current policy, published after the booking, allows refunds until 6 pm; it does not govern this case.

Decision. For property A, the house rule no longer allows a free change, but the property's disruption exception lets a duty manager waive the first night with a recorded reason. The valid action is to route to that role, not to promise the waiver. For property B, moving the stay depends on property A: if A shifts, B's first night overlaps the old A dates. The agent may request a date change on B but cannot waive the prepaid charge. The case is labeled needs approval at A and change request at B, with both linked cells cited.

Where review is needed. If the booking record lacks the rate plan code for property A, the rule is unresolved and the case goes to a reviewer. Applying property A's current 6 pm deadline would produce a wrong grant that looks generous and correct.

Time zone, rate plan or effective date: which assumption flips a hotel decision?

Time zone, rate plan identity and the policy effective date are the three assumptions that flip outcomes most often, so each should be recorded as a field rather than inferred. A refund deadline computed in the contact center's time zone instead of the property's can move a case across the line by hours.

Show uncertainty explicitly. Use a resolved, inferred or unresolved status for every rule a case relies on. Resolved means the rule came from a linked record. Inferred means a reviewer chose it from indirect evidence, such as a rate code pattern, and wrote down why. Unresolved means the case should be held out of training labels or used only to test whether a model asks for the missing record.

Channel is the fourth assumption. A booking made through a third-party site can carry that site's terms alongside the hotel's, and the reservation record may show only a channel code. Where the channel's terms are not linked, label the hotel-side decision and mark the channel-side question as out of scope for the case.

Checking coverage before the data trains or tests anything

Publish the coverage counts with the dataset: cases per property, per policy type and per version, plus the share of rules marked resolved, inferred and unresolved. A buyer can then see whether the set covers the properties and rules their model will face, or only the flagship property.

Split by property and by time. Holding out whole properties tests whether a model reads the rule from the record instead of memorizing one hotel's habits, and holding out a later period tests whether it handles a version change. Keep each disruption scenario's dependent bookings in the same split.

Grade the reservation state the model would leave, not the tone of its reply. The tau2-bench evaluation documentation describes this approach for agent benchmarks: reference actions are replayed to produce a target database end state, and any agent trajectory that produces an equivalent end state passes. For hotel cases the end state is the reservation record after the action, including dates, charges, waivers with approver role and any open request.

Getting multi-property records ready for licensing

Before policy histories and reservation cases leave the company, settle who controls them. A management company, an owner and a brand may each hold part of the record set under management, franchise or license agreements, and those agreements decide who can license what. Removing guest identifiers does not by itself create that right, so counsel should review the agreements for each property in scope.

Then prepare the records. Replace guest names, loyalty numbers and confirmation numbers with stable tokens, express stay dates as offsets where exact dates are not needed, generalize property identity if the buyer only needs segment and region, and keep the policy version and rate plan links intact, because those links are the value.

Hotel and management groups that want to know whether their reservation and policy histories could become a licensed dataset can start with SourceX's short company assessment, which needs no data upload. From there SourceX covers sourcing, rights review, contracting, preparation, delivery and payment, and nothing is delivered until the group has approved scope and terms.

Frequently asked questions

Why not label hotel cases with the brand's published policy?

Because properties and rate plans within one brand often differ from the brand default, and the reservation carries the terms the guest actually bought. Labels from the brand page are wrong exactly in the cases where a property or rate plan departs from it.

What is a property and policy version coverage matrix?

It is a table of policy types, each with what makes it vary, the date that anchors its version, the evidence a case must link and the legal or authority checks that can override it. Counting cases per property, policy type and version shows where a hotel dataset is thin.

Should a hotel dataset use the booking date or the stay date for policy versions?

Use the booking date for purchase-time rules such as cancellation, deposits and the total price, and the stay date for operational rules such as amenity hours. Record which anchor was used for each rule so reviewers can check it.

How do service animal rules fit into a pet policy dataset?

Treat them as an override, not as part of the pet policy. Federal guidance says hotels cannot charge pet deposits or fees for service animals or limit those guests to pet-friendly rooms, so cases should record when the override applied.

How should a model be graded on multi-property hotel cases?

Grade the reservation state it would leave, including dates, charges and any approved waiver, against the reviewed label. Test on held-out properties and later periods so the model must read rules from records instead of memorizing one hotel.

Sources

  • Google's Hotel Prices reference sets refundability per result, package or rate, with rate-level terms taking precedence and refund deadlines in hotel local time. Source
  • The schema.org vocabulary applies checkinTime and checkoutTime to both lodging businesses and individual lodging reservations. Source
  • The FTC's fees rule covers short-term lodging from May 12, 2025 and requires total prices that include mandatory fees. Source
  • DOJ guidance says hotels may not charge pet deposits or fees for service animals or limit those guests to pet-friendly rooms. Source
  • The ADA Title III regulations set reservation duties for places of lodging, including describing accessible features and holding and blocking accessible rooms. Source
  • The tau2-bench evaluation documentation grades agents on whether their database end state matches a target end state. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies