Revenue blog · Forecasting · 2027 planning guide
Can You Trust an AI Hotel Forecast in 2027? Seven Checks Before Acting
To judge AI hotel forecasting accuracy, test forecasts made with information available at the decision time, against a sensible simple benchmark, for the same target and horizon. A convincing explanation does not show how the forecast will perform. Neither does a close fit to the historical data used to build it.
This 2027 planning guide was prepared in November 2026. It sets out questions a hotel can use to evaluate a forecast; it does not predict how accurate future AI systems will become. The examples are hypothetical and are not a RevPerfect model evaluation.
1. Define the target and check the sources
Start with the question the forecast must answer. Rooms sold for a stay date, room revenue for a month and unconstrained demand are different targets. A hotel that sells its last available room has observed sales at its capacity limit; that result does not reveal every booking it might have accepted with unlimited inventory.
Ask for a short definition sheet: property, stay dates, hotel-local as-of time, available inventory, cancellations treatment, currency and revenue basis. For room revenue, agree whether tax, fees and package allocations are included. Confirm which source supplies each field and how late or missing files appear. A missing feed should prompt a completeness warning, not silently become zero demand.
If a number is supplied by the hotel's RMS or finance team, label it as that supplied forecast. Showing it beside actuals or explaining a variance does not mean the reporting platform generated it. Our hotel demand forecasting guide covers the wider operating process.
2. Preserve what was known at the historical cutoff
Choose a past decision time and reconstruct the information available then. For example, a forecast made 14 days before arrival must not use a cancellation received three days before arrival. Revised event attendance, final group pickup and corrected actual revenue also need their own known-at dates.
Ask the supplier to show the forecast version and its source cutoff. If the only available file is today's revised history, say that the historical test has limitations. Do not quietly treat it as the record the revenue manager saw at the time.
3. Test dates the method did not learn from
Hyndman and Athanasopoulos explain why forecast evaluation needs observations outside the fitting data. Their chapter also describes rolling-origin evaluation, which moves the historical cutoff forward while keeping future observations out of each forecast.
For a hotel review, ask to see ordinary weeks as well as difficult ones. Include relevant seasonal periods, event nights, cancellations and changes in inventory. List exclusions. A test made only on quiet dates cannot settle whether the method is useful on the dates where the hotel faces its largest decisions.
Keep a final review period separate from repeated model selection. Otherwise, each adjustment can become tailored to the same familiar answers. The appropriate duration depends on the hotel's history and decisions; this guide does not prescribe a universal number of months.
4. Compare the same horizon against a useful benchmark
A seven-day forecast and a 90-day forecast answer different questions. Request results by decision horizon, and give the benchmark exactly the same source cutoff, stay dates and target. A seasonal reference or the hotel's existing approved method can be a useful starting comparison; choose one that makes sense for that property.
Three nights: compare errors in rooms
Each row represents a separate seven-day-ahead forecast with the same information cutoff for both methods. Actual rooms sold are the later observed outcome.
| Night | Actual | Forecast | Benchmark | Absolute errors: forecast / benchmark |
|---|---|---|---|---|
| A | 80 | 85 | 78 | 5 / 2 |
| B | 90 | 88 | 84 | 2 / 6 |
| C | 100 | 92 | 90 | 8 / 10 |
Mean absolute error: forecast (5 + 2 + 8) / 3 = 5 rooms; benchmark (2 + 6 + 10) / 3 = 6 rooms. These three rows teach arithmetic. They establish no general model superiority.
Mean absolute error, or MAE, is the average size of the errors without their direction. The forecasting text also cautions that mean absolute percentage error, or MAPE, is undefined or infinite when an actual is zero and can become extreme near zero. Ask how zero-demand dates are handled before accepting an impressive percentage.
5. Look for bias and consequential mistakes
A single average can hide repeated overforecasting in one part of the business and underforecasting elsewhere. Keep the signed error beside the absolute error and define its direction. In the example, actual minus forecast is −5, +2 and +8 rooms. That tells a different story from “five rooms average error”.
Inspect relevant segments and events without slicing the sample into meaningless fragments. Ask whether cancelled groups, sellout dates or room closures explain large errors. Then connect the error to the decision: staffing, accepting a group and changing a commercial plan have different consequences. A universal “good accuracy” threshold ignores that context.
6. Review uncertainty and capacity together
Ask for a range and an explanation of what it represents, where the method supports one. How often did comparable historical outcomes fall within that range? What would the team do at its lower and upper ends? A narrow range is useful only when it is credible.
Keep capacity visible. A rooms-sold forecast should be interpreted against available rooms; an unconstrained-demand estimate may exceed capacity but must be labelled accordingly. An AI explanation should not erase a room closure, assume every enquiry becomes a booking or present an estimated event effect as a confirmed reservation.
7. Record overrides and review the result
Keep the original forecast, the revised value, the person approving the change and the reason. “New signed group” is more useful than “management adjustment”. Set a review date and compare both versions with the eventual outcome. This helps distinguish informed judgment from a habit of moving numbers toward a target.
Before using a forecast for a consequential decision, confirm: target defined; sources current; historical cutoff preserved; comparable benchmark available; errors examined; uncertainty understood; and an accountable reviewer named. If a critical item is missing, narrow the decision or request the evidence.
Forecast accuracy questions
What is a good accuracy percentage for an AI hotel forecast?
There is no universal percentage. Evaluate the same target and horizon against a suitable benchmark, inspect bias and important dates, and relate errors to the hotel's decisions. Percentage errors also need care when actual values are zero or small.
Does a forecast explanation prove the prediction is reliable?
No. Reliability requires an evaluation against later observed outcomes using information available at each historical decision time. A clear explanation helps review the assumptions but does not replace that evaluation.
Is this a test of a RevPerfect AI forecasting model?
No. The figures are hypothetical examples of forecast checking. RevPerfect provides reporting and guidance from supported hotel data; this article does not report a benchmark of a RevPerfect AI model.
Make forecast checks part of the reporting routine
For 2027, ask for reviewable evidence before adding more autonomy. See the technology answers for the product boundaries, and the reporting automation guide for keeping inputs, dates and exceptions clear in a repeatable report.