Seasonal and Ski Resort Hotels: Running a 200-Day Revenue Year with AI

A compressed season does not just shrink the calendar. It removes the margin for error from every pricing and staffing decision you make, and it concentrates a full year of consequences into about twenty weeks. Here is what the 2025-26 winter proved, and the operating model that absorbs the next one.

Every seasonal hotel operator I have worked with can tell you their break-even occupancy. Very few can tell you what happens to it when the season is three weeks shorter than planned.

The 2025-26 North American winter made that question concrete. US ski areas recorded an estimated 52.6 million snowsports visits, roughly nine million fewer than the prior season and 9.1 percent below the ten-year average. National average snowfall came in at 112 inches against a ten-year average of 169 — about a third below normal and the lowest total in more than a decade. Colorado, the largest single market in the country, posted its steepest annual decline in decades.

The lodging consequence followed directly. Western winter mountain destinations closed the season down 7 percent in occupancy year over year, with ADR up only 1.2 percent and full-season revenue down 5.9 percent. Colorado and Utah were worse than the regional average, running 6.7 percent below the prior winter on combined in-the-bank and on-the-books occupancy for November through April, against a 0.5 percent decline across the rest of the region.

None of that is a management failure. Nobody controls snowfall. What separates the properties that finished the season intact from the ones that did not is narrower and more interesting: how quickly they detected the shortfall, how they priced into it, and whether their labor plan could bend without breaking.

This piece is about that machinery. Not a forecast of next winter, which nobody can produce honestly, but the operating model a seasonal resort hotel needs so that the next bad winter costs it less — and the next good one earns it more.

The arithmetic of a compressed year

A year-round urban hotel earning 8 percent of annual GOP in a weak month has eleven more months to recover it. A seasonal resort earning 40 percent of annual GOP in a six-week window does not. That is the entire difference, and it propagates into every decision the property makes.

Three consequences follow, and they compound.

Error cost is concentrated. A two-point pricing error on a peak holiday week is worth more than the same error applied to an entire off-season quarter. Most revenue governance is calibrated the other way around: teams review pace weekly with equal attention regardless of the revenue weight of the period under review.

Fixed cost does not scale down with the season. Debt service, insurance, property tax, snow removal contracts, and the core management team persist for twelve months while revenue arrives in perhaps seven. The off-season is not a low-margin period, it is a cost floor you fund out of peak-season contribution.

Labor has to be built and dismantled annually. A seasonal property does not manage headcount, it manages a hiring and offboarding cycle, with lead times measured in months and, for international pipelines, in visa filing windows that close long before you know what the snow will do.

Put those together and the operating problem becomes clear. Seasonal resort hotels do not need better dashboards. They need earlier signal — enough lead time to move the levers that actually have lead time.

Look at the last two rows together, because they contain the most important strategic lesson of the season. Vail Resorts saw visits at its Colorado and Utah properties fall roughly 25 percent, and lift ticket revenue fall only 5.6 percent. That gap is not luck. It is the mechanical result of having sold the season before the season, transferring snowfall risk to the pass holder in the summer.

The lesson hotels should take from the pass model

The mega-pass has been analyzed to death as a lift product. It has been under-analyzed as a risk instrument, which is what it actually is, and which is what makes it relevant to lodging.

A season pass converts a weather-contingent transaction into a committed one, months in advance, at a known price. For 2026-27 the Epic Pass opens at $1,089 and the Ikon Pass at $1,349, with year-over-year increases moderating to post-COVID lows. Both companies are now competing hard for younger cohorts — Vail cut Epic and Epic Local pricing 20 percent for ages 13 to 30, and Ikon introduced a group product for buyers aged 23 to 28. That is a deliberate move to lock in lifetime-value cohorts early, and it tells you where the demand risk is perceived to sit.

Hotels generally cannot sell a pass. But the underlying mechanic — pre-commit demand at a known price, well before conditions are knowable — has direct lodging analogues that most seasonal properties use casually rather than systematically:

Multi-season loyalty rates for repeat winter guests, booked in the prior spring. Deposit-backed advance purchase inventory released in tranches during the summer. Corporate and group blocks for the peak weeks contracted twelve to eighteen months out. Owner-and-friends programs at resort properties with residential components. Committed room blocks tied to race weeks and events that are on the calendar a year ahead.

The AI contribution here is not the offer itself. It is knowing how much inventory to pre-commit, at what price, in which weeks. That is a straightforward constrained optimization once you can estimate the distribution of outcomes for a given week rather than a single point forecast — and a distribution is exactly what a conditional model produces and a same-time-last-year comparison does not.

Building a demand signal stack that actually leads

Most seasonal properties monitor two things: their own pace, and their comp set's rates. Both are lagging indicators. Pace tells you what already happened to bookings. Comp set rates tell you what your competitors already decided.

The signals that actually lead are external, public or cheaply licensed, and almost nobody in independent mountain lodging is systematically ingesting them. Modern RMS platforms increasingly consume this class of data, and the better systems will ingest hundreds of demand signals including weather, event calendars and flight search volume. The failure mode at seasonal properties is usually not the absence of a tool. It is that nobody has specified which signals matter for this property, at what lead time, moving which decision.

Two design points matter more than the specific list.

First, do not blend signals with different lead times into a single score. A road closure and a season pass sales trend are both demand information, but one should change tonight's overbooking decision and the other should change next season's group strategy. Collapsing them into one composite index produces a number that is not actionable at any horizon.

Second, weather is a conditioning variable, not a predictor by itself. Snowfall does not book rooms. Snowfall changes the conversion rate of demand that is already searching, and it changes it differently by segment — a drive-market weekend skier responds to a storm within 72 hours, a destination guest who bought airfare in October does not cancel over a thin base and generally shows up regardless. A model that treats those two segments identically will produce confident nonsense. Cornell's work on machine learning for hotel demand makes the general point well: segment-level structure carries most of the forecasting value, and aggregate models discard it.

Pricing the four seasons of a seasonal hotel

Seasonal resort pricing is usually described as peak and off-peak. That framing is the source of a great deal of lost revenue, because it collapses four genuinely different pricing problems into two buckets.

Peak, roughly 8 to 12 weeks. Demand exceeds supply on most nights. The job is not demand generation, it is mix optimization and length-of-stay control. Errors here are almost always errors of insufficient rate discipline or badly designed minimum-stay rules that fragment inventory.

Core season, roughly 10 to 14 weeks. Demand is present but responsive. This is where an RMS earns its money and where most properties are, in fact, reasonably competent.

Shoulder, roughly 8 to 12 weeks. The hardest problem in the seasonal calendar, and the least well handled. The data shows why: Mammoth Lakes in April 2026 ran a $369 ADR at 32.8 percent occupancy, delivering $121 RevPAR — a market whose rate held up while its occupancy fell through the floor. That is a market pricing shoulder as discounted peak and finding no takers.

Off-season. Either a genuine second season, or a cost-containment exercise. Both are legitimate; confusing one for the other is not.

The Beech Mountain pattern is the one worth studying, because it describes an asset with two real seasons rather than one season and a trough. Its summer rate holds at roughly 75 to 80 percent of its winter peak. That is not a discount, it is a different market being sold a different product at a defensible price.

The regional data suggests more properties should be pursuing that structure. After the poor winter, western mountain destinations saw summer bookings jump sharply, with full-summer occupancy up 4.9 percent and gains in every month but October. Summer at altitude is becoming a structurally stronger season, and a property still treating May through September as a caretaking period is leaving a second revenue year unbuilt.

The contribution floor, and why discounting the shoulder usually destroys value

The instinct when shoulder occupancy softens is to drop rate. In a seasonal asset this is frequently value-destructive, and the reason is a number most properties have never actually calculated: the contribution floor.

The contribution floor is the rate below which an occupied room fails to cover its own variable cost — housekeeping labor and supplies, utilities, commission and payment cost, plus the marginal cost of keeping a department staffed at all. At a full-service mountain resort with restaurants, spa and skier services, that last term is the one that dominates and the one that is almost always omitted. Keeping a restaurant open for a 30 percent-occupancy Tuesday has a cost that does not appear in a rooms-level flow-through calculation.

Resort-class properties carry materially higher labor per occupied room than any other segment, which raises the floor and narrows the window in which discounting makes sense. The practical implication is uncomfortable but clear: in the deep shoulder, a rate cut that fills rooms below the loaded floor moves money from the P&L to the guest and adds work for the team. Selectively closing floors, wings or outlets and holding rate is frequently the higher-contribution answer.

This is precisely the kind of calculation an AI layer is good at and a spreadsheet is bad at, because the floor is not a constant. It moves with occupancy, with which outlets are open, with day of week, and with the labor mix on shift. Modeling it as a curve rather than a number changes the decision on a meaningful share of shoulder nights.

Seasonal labor: the forecast is the hiring plan

Labor is where seasonal volatility becomes most expensive, because the decisions have the longest lead times and the least reversibility.

The 2026 environment is not forgiving. US hotels are projected to pay $131 billion in wages and benefits this year, up from $128 billion in 2025. Roughly 76 percent of hotels report operating short-staffed, with the sharpest gaps in housekeeping, front desk, culinary and maintenance, against annual turnover running 70 to 80 percent. Seasonal mountain properties layer on a further constraint: heavy reliance on J-1 and H-2B pipelines, whose filing windows and processing timelines have become an operational variable rather than an administrative detail. Specialist staffing firms have built entire businesses around this constraint.

You file for headcount in the spring for a winter whose conditions are unknowable. That is the structural problem, and it has no clean solution. But it has a much better and a much worse version, and the difference is whether your headcount request came from a demand model or from last year's number plus a fudge factor.

A workable approach has three parts. Build the ramp from the forecast distribution rather than the point estimate, sizing the committed core to something like the 30th-percentile season and the flexible tier to the upside. Stage arrivals against the actual occupancy curve rather than a single start date, which is where most seasonal properties burn payroll on underoccupied November weeks. And instrument the peak weeks for overtime, because peak-season overtime is where a good season quietly gives back its margin.

The column that repays attention is the last one. Note that the summer row is not a scaled-down winter — it is a different staffing shape, because the guest is different. Weddings, groups, hiking and mountain-bike traffic load F&B and grounds rather than skier services and bell. A property that staffs summer as a diluted winter will be simultaneously overstaffed and understaffed within the same week.

What to build, and in what order

The industry has moved past the question of whether to invest. Eighty-five percent of hotels expect to allocate at least 5 percent of IT budget to AI tools in 2026, and 82 percent expect AI usage to increase across their organization within the year. The question is sequencing, and at a seasonal property sequencing has a hard constraint that year-round properties do not face: you get one live test per year per season. A model deployed in January cannot be meaningfully retrained and re-validated until the following January.

That constraint should govern the roadmap. Build the things that can be validated against history first, and reserve the live-fire deployments for models you have already back-tested across several seasons.

A realistic first-year sequence looks like this.

Spring, before the labor filings. Build the conditional demand model and the labor ramp model together, in that order, because the second consumes the first. Back-test both against at least three prior seasons, including a bad one. If your history does not contain a bad season, weight 2025-26 heavily — you now have one.

Summer, into the shoulder. Stand up the contribution-floor engine and run it in shadow mode through the spring or fall shoulder. Compare its close-and-hold recommendations against what the property actually did, and reconcile the difference in dollars. This is the fastest credibility win available, because the shoulder is where instinct is least reliable and the counterfactual is easiest to compute.

Pre-season. Deploy pre-commit inventory tranching for the following year using the forecast distribution. Start conservatively — a modest share of peak-week inventory — and measure realized value against the unconstrained alternative.

Second year onward. Build the summer demand program properly. This is the highest-ceiling initiative on the list and the slowest to mature, because it requires actual market development rather than optimization of existing demand.

Properties starting this work usually discover the hard part is not the modeling. It is that three seasons of reservation history exist only as nightly occupancy roll-ups, labor actuals were never mapped to department and day, and nobody has ever written down the variable cost of opening the restaurant. A structured audit surfaces those gaps in weeks rather than months, and it is generally worth doing before committing to any vendor. If that is where you are, our AI Revenue Optimization & Forecasting engagement is built around exactly this sequence for seasonal and resort assets.

What the 2025-26 winter should have taught the industry

There is a temptation, after a season like the last one, to file it under bad luck and move on. That reading is too comfortable.

The resort operators who transferred snowfall risk to pass holders in July lost 5.6 percent of lift revenue on a 25 percent visit decline. The lodging properties who sold into January conditions lost 5.9 percent of revenue on a 7 percent occupancy decline — and did so while holding rates that were already deterring price-sensitive winter visitors. The two outcomes look similar in percentage terms. They are not similar in kind. One was a designed risk position. The other was an accident of timing.

Meanwhile, ski areas reinvested roughly $569 million in capital during the same season — 45 new lifts and 52 upgrades, about $22 per skier visit. The mountains are still being built. The demand-side instrumentation at the lodging properties that serve them is, in most cases, a decade behind.

That gap is the opportunity. Not because AI is novel — seasonal pricing strategy and demand forecasting are well-trodden ground — but because the specific combination of conditional forecasting, contribution-floor discipline and forecast-driven labor staging remains rare in independent mountain lodging, and the properties that build it will be structurally better positioned the next time the snow does not come.

It will not come again. That is the only forecast in this piece I will make with confidence.

Frequently asked questions

Why do standard revenue management systems underperform at seasonal resort hotels?

Most RMS engines learn from same-time-last-year patterns and assume the shape of demand repeats. At a seasonal resort the shape is set by conditions that do not repeat: snowfall, road access, event calendars and pass-holder behavior. In 2025-26 national snowfall came in at 112 inches against a ten-year average of 169, and US skier visits fell to 52.6 million from 61.5 million. A model anchored on last year's curve read that entire season as a pace failure and recommended discounting into a demand collapse it could not see. Seasonal assets need conditional forecasting driven by external signals, not calendar-anchored pattern matching — which is a different model specification, not necessarily a different vendor.

What external data signals actually predict ski resort hotel demand?

In rough order of forecasting value: base depth and snowfall forecasts for short-lead transient; season pass sales and pass-holder redemption history for the committed base; airline seat capacity and fares into the nearest gateway for destination mix; origin-market search volume for pricing by feeder market; the event, race and school calendar for compression nights; and state DOT road and pass status for same-day cancellations. Each carries a distinct lead time — from 0-3 days for road hazard to 90-365 days for events — so they belong in different parts of the decision stack rather than blended into a single composite index that is actionable at no horizon.

How should a seasonal hotel price the shoulder season?

Not as a discounted version of peak. Shoulder demand comes from a different guest with different elasticity, and the binding constraint is the contribution floor — the rate below which an occupied room fails to cover its variable service cost plus the incremental labor required to keep the relevant departments open. The mountain data shows how steep the swing is: Mammoth Lakes ran a $369 ADR at 32.8 percent occupancy in April 2026, producing $121 RevPAR. The right posture is to compute a per-room contribution floor by department and day, price above it, and use targeted segment offers rather than an across-the-board cut that trains the market to wait.

Can AI help with seasonal labor planning, or is that just a hiring problem?

It is both, and the forecasting half is the tractable half. US hotels are projected to pay $131 billion in wages and benefits in 2026, 76 percent of hotels report operating short-staffed, and seasonal properties additionally depend on J-1 and H-2B pipelines with filing windows that close months before conditions are knowable. AI does not fix visa policy. But a demand model that produces a defensible headcount curve by department and week lets you file for the right number, stage arrivals against the real ramp instead of a single start date, and cap overtime in the peak weeks where margin quietly leaks. The forecast is the input to the hiring plan, which is why forecast error at a seasonal property is expensive twice.

Is AI worth deploying at a single seasonal property with a 200-day operating year?

Yes, and arguably more so than at a year-round asset, because error cost is concentrated. A property earning most of its annual GOP in roughly twenty peak weeks has far less time to correct a bad pricing or staffing call. The two highest-return builds are a conditional demand model that ingests weather and pass signals, and a labor ramp model tied to that forecast; both run largely on data the property already holds plus free public feeds. The real prerequisite is not scale but data hygiene — multi-season reservation history at stay level, and labor actuals mapped to department and day. Verify you have both before scoping anything else, because that single check changes seasonal AI project timelines more than the choice of model ever does.