How Accurate Is Your Forecast? Benchmarking MAPE Across Hotel Departments

Every hotel produces a forecast and almost none of them grade it. This is the owner's guide to scoring the forecast the way you score RevPAR: which error metric to use, what an acceptable error band looks like by department and horizon, how to catch the bias that sank the 2025 budget cycle, and how to put a dollar figure on every point of error before it lands in the labor line.

The forecast nobody grades

Every hotel produces a forecast. Almost none of them score it. The revenue manager submits a 90-day rooms forecast on the first of the month, the director of finance rolls it into a P&L reforecast, the executive housekeeper turns it into a schedule, the executive chef turns it into a purchase order, and the general manager presents it to ownership. Thirty days later the actuals arrive, everyone looks at the variance to budget, and the forecast itself quietly disappears. Nobody asks how far off it was, whether it was off in the same direction it was off last month, or which department paid for the miss.

That is the gap this article is about. Forecast accuracy is the one operating metric that touches labor, purchasing, pricing, and owner credibility simultaneously, and it is the one metric almost no independent hotel tracks with any discipline. The 2025 budget cycle made the cost of that visible across the whole industry. According to HotStats data published by HotelData.com, US rooms revenue finished the third quarter of 2025 13.2% below budget and 11.9% below budget year to date. Even after hotels reforecast mid-year, rooms revenue still trailed the revised forecast by 4.7% in the quarter and 5.2% year to date. The forecasts were optimistic in November, still optimistic in June, and nobody had a mechanism to catch the bias until the money was gone.

Meanwhile the labor line, which is the largest controllable cost in the building and the one most directly driven by the forecast, kept rising. Hotel Management reports labor cost per occupied room at $48.32 and climbing at double-digit rates at many full-service properties, which is why a 5% scheduling error at a 300-key hotel is a six-figure problem. The 2025 Hotel Labor Costs & Trends Report found that operators protected margin in 2025 not by cutting teams but by cutting hours per occupied room 7% to 15% in guest services, housekeeping, and management. You cannot cut hours to demand you did not forecast correctly. Precision in labor is downstream of precision in forecasting, and precision in forecasting starts with measuring the error.

What follows is a practical framework for scoring the forecast: which error metric to use and when, what an acceptable error band looks like by department and by horizon, how to detect the systematic bias that MAPE hides, and how to translate a point of forecast error into dollars of labor, food, and revenue so the whole executive committee understands what accuracy is worth. It ends with a 90-day implementation plan that a revenue manager and a controller can run without buying anything.

A forecast that is never scored is not a forecast. It is a guess with a spreadsheet around it. The moment you start grading it, it starts getting better, because the people who make it finally know what better means.

Why hotels do not score their forecasts

The reasons are structural, not personal. First, the forecast has too many owners. Rooms belongs to revenue management, F&B covers belong to the outlet managers, banquet revenue belongs to catering sales, and the consolidated P&L forecast belongs to finance. Each group produces its number on its own cadence, in its own spreadsheet, using its own definition of the forecast date. There is no single version to score, so nobody scores any of them.

Second, the forecast is overwritten. The 90-day forecast submitted on September 1 for October is replaced by the one submitted on October 1, which is replaced by the 10-day forecast, which is replaced by the day-of pickup. By the time October actuals land, the original September 1 view no longer exists in the system. You cannot measure the error of a number you did not keep. This is the single most common reason forecast accuracy programs fail on day one: the property has no archive of what it said it would do.

Third, the metric is politically uncomfortable. A revenue manager who is scored on RevPAR performance has every incentive to forecast conservatively so that actuals beat the number, and a director of sales who is scored on booking pace has every incentive to forecast group aggressively so that the pipeline looks healthy. Both are rational, and both introduce bias that compounds when finance sums them. Scoring the forecast exposes those biases, which is precisely why it is valuable and precisely why it is resisted.

Fourth, most hotels have never seen what good looks like. Nobody has told the GM that a 30-day rooms forecast at 8% MAPE is excellent and 20% is a problem, so there is no target to manage to. The rest of this article fixes that.

Choosing the right error metric

MAPE, mean absolute percentage error, is the default for a reason. It is intuitive, it is unit-free, and a GM can understand "we were off by 12% on average" without a statistics lesson. Calculate it by taking the absolute difference between forecast and actual for each day, dividing by the actual, averaging across the period, and expressing the result as a percentage. A 30-day forecast that predicts 220 rooms on a night that actualizes at 200 contributes a 10% error for that day.

But MAPE has well-documented failure modes that matter in a hotel. The Institute of Business Forecasting and practitioners like Demand Planning LLC point out that MAPE explodes on low-volume days, because dividing a small absolute error by a small actual produces a huge percentage. A resort forecasting 6 banquet covers on a Tuesday that actualizes at 3 has a 100% error on that day, which will swamp the month. MAPE also treats over-forecasting and under-forecasting asymmetrically: you cannot be more than 100% under, but you can be infinitely over. And, most important for a hotel, a forecast can carry a respectable MAPE while being consistently wrong in the same direction, because the absolute value strips the sign.

The fix is not to abandon MAPE but to pair it with two companions. The first is weighted MAPE, usually written WMAPE or WAPE, which sums the absolute errors and divides by the sum of actuals. It weights busy days more heavily than slow days, which is what you want: being off by 20 rooms on a 300-room sellout night matters more than being off by 2 rooms on a 30-room night. WFM Labs recommends WAPE as the primary accuracy metric for workforce planning for exactly this reason. The second companion is bias, sometimes called mean percentage error, which is the same calculation as MAPE without the absolute value. Bias tells you the direction of the error. A forecast with 12% MAPE and +9% bias is over-forecasting almost every day. That is a much more actionable finding than 12% MAPE alone, and it is the finding that explains the 2025 budget season.

The practical recommendation for a full-service hotel is simple. Report MAPE and bias for rooms at every horizon, report WMAPE and bias for every labor-driving department, and use MAE in units for any outlet or revenue center that regularly sees single-digit daily volume. Present all of them on one page. The point is not statistical purity. The point is that the executive committee sees the same scorecard every month and knows which number to react to.

Accuracy benchmarks by department and horizon

The honest answer to "what MAPE should we hit" is that it depends on the horizon, the department, and the demand pattern of the property. But an answer that depends on everything is useless to an owner, so here are the bands that hold across enough properties to be a fair starting point. Cross-industry guidance summarized by Hospitality Net and Lighthouse puts a MAPE under 10% as excellent, 10% to 20% as acceptable, and 15% to 25% as the realistic range for seasonal or event-driven hotels. A large city-center hotel with steady corporate demand can regularly run 10% to 12% on a 30-day rooms forecast; a 50-room independent in a ski town may find 20% is as good as it gets in shoulder months. Cloudbeds makes the same point from the demand-surface side: accuracy should be measured against the horizon at which the decision is made, not at a single arbitrary point.

Academic benchmarks are tighter than operating benchmarks because they measure the model, not the whole process. The foundational Weatherford and Kimes comparison of forecasting methods, run on Choice and Marriott data, found pickup methods, exponential smoothing, moving averages, and regression to be the most robust, and hotels still using a simple pickup model have not been proven wrong by two decades of subsequent research. More recent work published in Tourism Management shows ensemble models reaching MAPE below 5.1% on daily occupancy and cutting error more than 80% against a naive benchmark, and the Boston Hospitality Review reports similar single-digit results for well-specified daily models. Those numbers are achievable. They are also achieved by a model with clean data and no committee editing its output. The operating benchmarks below are for the forecast that actually reaches the schedule.

Three things about that table deserve emphasis. First, the 7-day rooms number is the one that drives the housekeeping schedule, and it should be under 5% at almost any hotel because most of the business is already on the books; a 7-day rooms MAPE above 8% is almost always a data hygiene problem, not a demand problem. Second, restaurant covers are the hardest line in the building to forecast because they depend on in-house capture, local walk-in, weather, and competitive openings, none of which the PMS sees. That is why they are the first place to deploy a model with external inputs. Third, banquet definite business should be nearly perfect a week out, and if it is not, the problem is the BEO process, not the forecast.

Detecting bias: the error that MAPE hides

Return to the 2025 numbers. Rooms revenue was 11.9% below budget year to date, and the industry-wide reforecast was still 5.2% too high. Those are not random errors. A random error would be over some months and under others and would average close to zero. A forecast that is over by 5% to 13% every month for three quarters has a bias, and the bias has a cause: an assumption baked in during budget season, in this case a double-digit RevPAR growth expectation carried into a year that actually produced the first non-recessionary RevPAR decline on record, as reported by Hotel Online from the STR and Tourism Economics forecast. The assumption was never revisited because nobody was measuring the direction of the error.

The tool for catching this is the tracking signal, borrowed from supply chain control charts and explained well by Arkieva and SupliiChain. Sum the signed errors over the last several periods, divide by the mean absolute deviation, and watch the result. A forecast with no bias hovers near zero. When the tracking signal moves beyond roughly plus or minus 4, the forecast has drifted out of control and needs to be recalibrated. The beauty of the tracking signal is that it catches a small consistent bias long before it becomes a large one. A forecast that runs 3% high every month has a MAPE that looks fine and a tracking signal that climbs steadily until someone has to explain it.

In a hotel, bias almost always has a human signature. Revenue managers under-forecast so that they beat the number. Sales directors over-forecast group so that the pipeline looks full. Outlet managers over-forecast covers so that they are staffed for the good night. Finance smooths everything toward budget so that the owner call is calm. None of these people are doing anything wrong by their own scorecard. The forecast accuracy program exists to give them a shared scorecard, and the bias column is where the conversation starts.

Translating error into dollars

The reason forecast accuracy is worth an owner's attention is that every point of error has a cost, and the cost lands in three places: labor scheduled for demand that did not arrive, food purchased for covers that did not show, and revenue left on the table when the forecast was low and the hotel priced or staffed as if demand was soft. The labor cost is the largest and the easiest to compute. The collaborative study on staffing demand forecasting in the hotel industry and Deputy's guidance on data-driven hotel staffing both make the same point: labor is scheduled against the forecast, not the actual, so forecast error converts almost directly into unproductive hours when the forecast is high and into overtime and service failure when it is low. The 2026 workforce management outlook from Hotel Online notes that AI scheduling tools have lifted forecast accuracy from the mid-50s to the high 80s in percentage terms and cut overtime 20% to 40% at adopting properties, which is a direct measure of what the error was costing before.

The food cost is smaller per point but compounds fast in a hotel with multiple outlets and banquets. The Champions 12.3 study of hotel kitchens found that every dollar invested in reducing kitchen food waste returned an average of seven dollars in operating savings, and that participating properties cut food waste by at least 10% and in some cases lowered food cost by 3 points or more. A case study of hotel food service waste published in Sustainability found that the majority of waste occurs at the serving stage, in buffets and plate waste, which is exactly the waste that a better cover forecast prevents, because production is set the day before against the forecast. Orbisk reports that hotels achieving 30% to 53% waste reductions all measure by outlet and meal period, which is the same granularity a good cover forecast requires.

Here is the worked translation for a hypothetical 250-room full-service hotel at 70% occupancy, $220 ADR, one restaurant, one bar, and a modest banquet operation, using the labor CPOR from Hotel Management and typical food cost and waste figures from the studies above. The point of the exercise is not the exact dollars, which will differ at every property. The point is that a controller can rebuild this table for their own hotel in an afternoon and put a number on the accuracy program before the first meeting.

Summed, a 5-point forecast error at this hotel carries somewhere between $290,000 and $425,000 of annual exposure before the credibility line, and a disciplined accuracy program plausibly recovers half of it. That is a return worth chasing with a spreadsheet and a monthly meeting, which is all the first phase requires.

Owners do not fire general managers for missing the forecast. They fire them for missing it in the same direction three quarters in a row and not knowing why. Bias is the number that ends careers, and it is the one number almost nobody reports.

Where AI actually helps, and where it does not

The forecasting vendor market has been loud about machine learning for a decade, and some of the claims are true. McKinsey's widely cited estimate, summarized by ToolsGroup and ThroughPut, is that AI-driven forecasting reduces error 20% to 50% against conventional methods, and the mechanism is the number of inputs. A spreadsheet pickup model uses three to five variables; a gradient-boosted or ensemble model can ingest 20 to 40, including weather, local events, competitor rates, airline capacity, search trends, and cancellation velocity. In hospitality specifically, the systematic review of deep learning for resort demand forecasting and the extended additive pickup work for SME hotels both show meaningful error reduction from richer inputs, with the important caveat that the gains are concentrated at the 30-day and 90-day horizons where the booking curve is still forming. At 7 days, most of the business is on the books and a good pickup model is already close to the ceiling.

The part vendors say less about is that a model cannot fix a process. Steve Morlidge's study of eight companies, published in Foresight, found that 52% of their forecasts were worse than a naive random walk, meaning that all the human adjustment, committee review, and executive override layered on top of the statistical forecast made it worse more often than better. Forecast value added analysis, laid out step by step in SAS's white paper and the IBF series on FVA, measures each step of the process against the step before it and asks a blunt question: did this person or this meeting make the forecast more accurate, or less? Hotels that run FVA typically discover that the revenue manager's statistical forecast is the most accurate number in the building and that every subsequent edit, from the DOSM's group adjustment to the GM's gut check to finance's smoothing, degrades it. That is not an argument against human judgment. It is an argument for measuring it.

So the honest sequencing is this. First, archive and score the forecast you already have. Second, run FVA to find out which steps of your process help and which hurt, and remove the ones that hurt. Third, and only third, replace the statistical core with a model that uses more inputs, because by then you will have a clean baseline to prove it against. Hotels that skip to step three buy a better engine and bolt it to the same leaky process, and then conclude that AI does not work. Properties that want help with the modeling layer often find it useful to start with a structured forecasting assessment rather than a software purchase; that is the work our AI Revenue Optimization & Forecasting engagement is built around, and the first deliverable is always the accuracy baseline described in this article.

A 90-day implementation plan

None of this requires new software. It requires a controller, a revenue manager, a shared spreadsheet, and a standing 30-minute meeting. The plan below is the one we run with independent hotels, and the milestones are deliberately unambitious in the first month because the archive is the hard part and everything else depends on it.

Two operating rules make this stick. The first is that the forecast is locked at submission. Finance can produce a separate reforecast, sales can produce a separate pipeline view, and the GM can annotate all of it, but the number the revenue manager submitted on the first of the month is the number that gets scored, and it does not change. The second is that accuracy and bias go on the revenue manager's and the department heads' scorecards with a weight that matters. The retail planning literature is unambiguous that planners score on what they are measured on, and hotel forecast owners are no different. Score them on beating forecast and they will sandbag. Score them on accuracy and they will forecast.

What the scorecard should look like

The monthly scorecard fits on one page and answers four questions in order. How accurate was the forecast at each horizon that drives a decision? Was it biased, and in which direction? Which step in the process added or destroyed accuracy? What did the error cost? A version we use lists rooms occupied, rooms revenue, restaurant covers by meal period, banquet covers, and total labor hours down the left side; shows 7-day, 30-day, and 90-day MAPE or WMAPE with the bias in parentheses across the top; colors any cell outside the action threshold from the benchmark table; and closes with the tracking signal trend and the dollar translation. It takes a controller two hours a month once the archive exists.

The scorecard changes the conversation at the executive committee. Instead of "why did we miss budget," the question becomes "the 30-day rooms forecast has run 6% high for four months, the tracking signal is at 5.1, and it is costing us roughly $14,000 a month in housekeeping hours; what assumption are we carrying that is wrong?" That is a question with an answer, and it is a question that makes the revenue manager, the controller, and the department heads allies rather than suspects.

Hotels that have run this for a year report the same three outcomes. Rooms MAPE at 30 days drops into the low teens or better within two quarters, mostly from removing bias rather than from any modeling change. Hours per occupied room falls because the schedule is built from a number people trust. And the owner call gets shorter, because the general manager arrives with a forecast that has a track record instead of a forecast that has a story. Forecast accuracy is not a revenue management metric. It is the operating discipline that makes every other metric in the building believable.

Frequently Asked Questions