What is a good MAPE for hotel occupancy forecasting?

Published benchmarks vary by market volatility and horizon, but single-digit MAPE (under 10%) on 30-90 day horizons is generally considered strong for property-level occupancy. Always ask on what data and horizon a MAPE was measured — a number without context is marketing, not evidence. Nexorev\u2019s 9.8% MAPE, for example, comes from public North Italy data backtests, not live hotels.

What data does a hotel demand forecast need?

At minimum: two years of reservation history with booking dates (to build booking curves), current on-the-books position, room inventory, and a calendar of events and seasons. Better systems add competitor rates, market demand indices, weather, and macro signals as refining inputs.

Why do hotel forecasts fail?

The usual causes: structural breaks the history cannot see (a new competitor, renovation, pandemic-style shocks), one-off events treated as recurring, data quality problems in the PMS export, and overfitting to noise in small properties where a single group booking distorts the pattern.

Hotel Demand Forecasting Explained: How Occupancy Predictions Actually Work

Why Forecasting Is the Whole Game

Every revenue management decision — tonight's rate, that festival weekend's minimum stay, whether to accept the 15-room group in May — is a bet on future demand. The forecast is the bet's foundation. A pricing engine bolted onto a bad forecast just makes wrong decisions faster. This article explains, in plain language, how hotel demand forecasting actually works, what the accuracy numbers mean, and how to interrogate any vendor's claims — including Nexorev's own.

The Core Object: the Booking Curve

For any future stay date, plot rooms on the books against days remaining until arrival. That is the booking curve, and it is remarkably stable per market and segment: a leisure Saturday in a North Italy lake town fills along a recognisable trajectory — slowly from 90 days out, accelerating inside 30, spiking in the final week. Forecasting, at its heart, means answering: is this date ahead of or behind its typical curve, and by how much? "Pickup" — the bookings still to come between today and arrival — is what the model actually predicts.

The Methods Ladder, from Naive to ML

Same-time-last-year (STLY): assume this date behaves like last year's equivalent. The baseline every fancier method must beat — and in stable markets, embarrassingly hard to beat by much.
Classical pickup models: additive or multiplicative pickup from historical curves, per day-of-week and season. The workhorse of hotel forecasting since the 1990s; transparent and robust.
Regression models: pickup plus explicit features — lead time, day of week, events, holidays, price level — with interpretable coefficients.
Machine learning: gradient-boosted trees and neural approaches (including LSTM sequence models) that capture non-linear interactions: how an event shifts the curve differently in shoulder season, how weather interacts with lead time. Gains are real but incremental — typically worth 1-3 MAPE points over well-tuned classical methods, not miracles.
Ensembles: blending several models, weighted by recent accuracy per horizon. Most production-grade systems land here, because different methods win at different lead times.

A deeper model-by-model comparison is in Forecasting Models Compared 2026.

The Inputs That Matter (Ranked Honestly)

On-the-books position and pace — dominant. Nothing predicts demand like demand already arriving.
Seasonality, day-of-week, holidays — the calendar explains most of the remaining variance in leisure markets.
Events — high impact, must be explicitly encoded; history alone misdates movable fairs and one-off concerts.
Competitor rates and market demand indices — context that refines the picture, especially for price-response estimation.
Weather, macro indicators — real but second-order; they trim errors at short horizons.

Reading Accuracy Claims: MAPE and RMSE Without the Fog

MAPE (mean absolute percentage error): on average, how far off the forecast is, in percent. A 9.8% MAPE on occupancy means a date forecast at 70% occupancy typically lands within about ±7 points of it. RMSE (root mean squared error) penalises large misses more; quoted in percentage points, it tells you about tail behaviour — a low MAPE with a high RMSE means the model is usually close but occasionally badly wrong, which for pricing is dangerous.

Three questions expose any accuracy claim:

Measured on what data? Real hotels, or backtests? In-sample (the model saw the answers) or out-of-sample?
At what horizon? Forecasting tomorrow is easy; 60 days out is not. A blended single number hides this.
Against what baseline? If STLY scores 12% MAPE in that market, a model's 9.8% is a modest, honest gain — not magic.

Worked example — Nexorev's own numbers held to this standard: the 9.8% occupancy MAPE and 6.4pp RMSE published on this site come from out-of-sample backtests on public North Italy tourism and market data with synthetic PMS-like booking curves — not from live hotels, because Nexorev is pilot-stage with no production deployments. They beat the static baselines in the same fixture (the simulated RevPAR gain vs static rules is +7.6%), and they will be re-reported per property, per horizon, when real pilots produce real data. That is the level of disclosure worth demanding from every vendor.

Why Forecasts Fail

Structural breaks: a new competitor opens, you renovate, a flight route closes — history stops being representative and models need explicit correction.
Event mishandling: movable fairs, biennial events, and one-offs pollute year-over-year patterns unless encoded.
Small-property noise: at 25 rooms, one group booking is a 40% occupancy swing; good systems forecast distributions, not just points.
Dirty inputs: PMS exports with test bookings, unlogged channel closures, or timezone-shifted timestamps quietly poison everything downstream.

From Forecast to Price

The forecast is half the machine. The other half is price response: how demand for this date shifts if the rate moves EUR 10. The RMS combines both — expected demand at candidate prices, times rate, minus channel costs — and recommends the revenue-maximising point within your floors and ceilings. This is why forecast transparency matters operationally: staff who can see why (pace ahead, event nearby) trust and apply recommendations; black boxes get overridden into irrelevance.

Next Steps

Nexorev's public backtest metrics — the full accuracy disclosure.
The live demo — see forecasts and rate reasoning on real market data.
15-minute founder call or contact form — methodology questions welcome.

Frequently Asked Questions

What is hotel demand forecasting?

Predicting rooms sold (and demand at each price level) for every future date from pace, history, seasonality, events, and market signals — the foundation of all RMS pricing decisions.

What is a good occupancy MAPE?

Under 10% at 30-90 day horizons is generally strong at property level — but only meaningful with the data, horizon, and baseline disclosed. Demand all three from any vendor.

What data is required?

Two years of reservation history with booking dates, current on-the-books, inventory, and an event calendar; competitor and market data refine further.

Why do forecasts fail?

Structural breaks, mishandled events, small-property noise, and dirty PMS data — more often than model weakness.