A hotel in Valencia knows tomorrow's room occupancy, but the breakfast team still works from a spreadsheet and yesterday's intuition. A sports group may arrive late, several room-only bookings may buy breakfast on the day, and departing guests may cluster around an early transfer. The useful forecast is not simply occupied rooms. It is the workload and purchasing demand that follow from those bookings.
A forecasting project should begin with a decision, its deadline and the cost of being wrong. For Spanish hotel groups, that might mean breakfast purchasing, housekeeping capacity or reception coverage. A model becomes useful when it changes one of those decisions reliably enough to justify its integration and ongoing operation.
Foundation models open options, but your hotel still needs evidence
Google Research introduced TimesFM-3 in August 2026, describing a forecasting model that can use related series and external features. This makes multivariate forecasting a timely option to evaluate. It does not demonstrate that a particular model will improve a Spanish hotel's purchasing or staffing decisions.
Deployment terms also matter. The official TimesFM repository distinguishes model versions and their licensing. At this review, its notice restricts downloaded TimesFM 3.0 weights to non-commercial, non-production use while permitting commercial use through authorised Google Cloud services under their terms. Check the licence of the exact weights and the terms of any hosted service before selecting a commercial implementation. A working research notebook is not sufficient evidence that a model can be deployed under the intended operating arrangement.
Treat a foundation model as one candidate alongside simpler methods. A seasonal baseline may perform well for a predictable breakfast service and be easier to maintain. The buying question is which approach makes better decisions on your data at an acceptable operating cost, with failure modes the team understands.
Specify the output in operational language
For breakfast purchasing, the target might be expected covers by service period, with a useful uncertainty range. For housekeeping, occupied rooms alone are insufficient: departures, stayovers, room types and planned maintenance produce different workloads. Reception demand may depend on arrival groups rather than total nights sold.
Choose one target for the first release. State who receives the output, when it arrives and what they can change in response. If the purchasing order closes at noon, a forecast calculated from data available at midnight is not the forecast that actually informed that order.
Keep the recommendation separate from an automatic commitment. A kitchen manager may adjust the proposed quantity because of a supplier constraint or an event absent from the booking system. Record the adjustment and its reason. Those observations can improve the process without pretending that every human override is either correct or a model failure.
Reconstruct what the team knew at the time
Forecasting tests can look impressive when they accidentally use information that was unavailable at the decision point. Final occupancy, actual weather and completed cancellations are easy to obtain later. They should not quietly replace the booking snapshot and forecasts that staff would have had earlier.
Store time-stamped booking snapshots, changes and source refresh times. Separate confirmed reservations from provisional groups. Mark missing or delayed feeds instead of filling them with values that look authoritative. When external information is used, preserve the version available at the moment the recommendation was made.
Forecasting: Principles and Practice explains time-series cross-validation, simple reference methods and prediction intervals. Apply those principles through repeated historical decision dates: train or provide context using the past, forecast the next period and compare the result with what happened. Keep the final evaluation period separate from model tuning.
Judge expensive mistakes differently
Overestimating croissant demand creates a different cost from underestimating the people needed to serve guests safely and promptly. A single average percentage error can conceal both. Agree measures appropriate to the target, including error at busy periods and the business effect of systematic underprediction.
Uncertainty should influence the recommendation. When booking patterns are unusual or a source is stale, the interface should show that confidence is weaker and invite review. A narrow-looking range from a model is not enough: check whether historical outcomes actually fell inside its ranges at the expected frequency.
Include difficult weeks in the evaluation. A refurbishment, an unusual group booking or a sudden cancellation can challenge assumptions that ordinary weekdays never expose. Show performance by relevant operating situation, while avoiding conclusions from tiny samples. A good pilot report makes limitations visible alongside the cases that worked.
Put a small decision service behind the dashboard
The software needs more than a chart. It needs scheduled data collection, validation, versioned calculations, a review screen and a record of the final decision. If a job fails, staff should know whether they are seeing yesterday's forecast or a current result. A documented fallback keeps operations moving.
For a first pilot, select a small group of comparable properties and one decision horizon. Run the forecast alongside the existing process before letting it influence orders. Record the original staff plan, the model recommendation, the final choice and the observed outcome. This helps distinguish model quality from the effect of operational changes.
The implementation contract should identify who owns the connectors, how model or vendor changes are evaluated and how customer data is protected. Require an export of input snapshots and evaluation results so the business can review the system without depending on an unexplained score in a supplier portal.
Our guides to cloud operating costs and software integration scope address two common sources of surprises in forecasting projects. Start with a decision valuable enough to justify both.
TuniCyberLabs builds AI-enabled operational software. Tell us which hotel planning decision still depends on a fragile spreadsheet. We can scope the data preparation, evaluation and workflow needed to test a useful alternative.