Seasonal Demand Forecasting: Building a Staffing Model That Survives Shoulder Season

Labor is the largest cost a hospitality operator can actually control, and in a seasonal market it is also the one most often set by guesswork. Seasonal demand forecasting means predicting how much business will walk through the door, day by day and daypart by daypart, far enough ahead to staff for it. Most operators still do this by copying last year's schedule and adjusting by feel, which works until the year stops resembling last year.

The room for that has narrowed. The National Restaurant Association reports that in 2024, salaries and wages including benefits ran a median of 36.5% of sales at full-service restaurants, against roughly 33% across 2010, 2013, and 2016. Its 2026 forecast puts real sales growth at 1.3% and reports that 42% of operators said their restaurant was not profitable in the prior year. When a third of revenue goes to labor and real growth is roughly one percent, a schedule that runs 8% heavy is most of the profit.

Seasonality widens that error rather than averaging it out. In the Myrtle Beach-Conway-North Myrtle Beach metro area, leisure and hospitality employment ran 32,300 in January 2026 and 45,600 in July, a 41% swing in six months, per Bureau of Labor Statistics data via FRED. A payroll that adds 13,300 jobs and then gives them back cannot be run on intuition, and the same holds anywhere customers arrive on a calendar.

Table of Contents

1. What Seasonal Demand Forecasting Has to Produce
2. Build the Forecast in Four Layers
3. What Better Forecasting Is Worth: Worked Math for a 90-Seat Restaurant
4. Where AI Helps and What It Must Not Decide
5. Four Failure Modes and the Early Signals
6. A 90-Day Build Before the Next Season
7. Frequently Asked Questions
8. Ready to Forecast Next Season Instead of Reacting to It?

What Seasonal Demand Forecasting Has to Produce

A seasonal demand forecast is useful only if it outputs the unit a manager schedules against, at the interval they schedule in, early enough to act on. For most operators that means covers or transactions by daypart, two to three weeks out, not monthly revenue.

That rules out most of what passes for forecasting. A projection in dollars tells a kitchen manager nothing: two Saturdays with identical sales need different staffing if one is four large parties and the other sixty two-tops. Forecast the driver of work, then convert to dollars. Three outputs are worth building, in this order:

1. Volume by daypart. Covers, check-ins, tickets, or jobs dispatched, for each service period of each day.
2. Required hours by role. Volume divided by a stated productivity standard: covers per server hour, rooms per attendant shift, calls per dispatcher. Write the standard down; most operators never have.
3. A confidence range. A single number invites false precision. "Thursday: 210 covers, likely 185 to 245" lets a manager schedule the 185 and put two people on call.

The range is the output almost everyone skips, and it is where the money is. Scheduling to a point estimate makes managers hedge upward on every shift, which is how a forecast that is accurate on average still runs a heavy payroll.

Last year's same week is a fair starting point and a poor answer, because the shape of a season moves even when its total does not. The Myrtle Beach Area Chamber of Commerce reported summer 2025 lodging occupancy flat against 2024 at about 70%, but hotel occupancy inside that average ran 71.1% in June, 76% in July, and 64.4% in August, with June down 4.8% year over year while July and August were up. Staff 2025 off 2024 and you were heavy in June, short in August, and finished with a season number that looked fine.

Build the Forecast in Four Layers

Build the forecast in four stacked layers, each answering a question the layer below cannot. Skip one and you get a model that is accurate in March and useless on the Fourth of July.

Layer 1: the baseline. Two to three years of daily transaction history, cleaned before anything else touches it. Remove closure days, mark remodels and menu changes, and flag one-off events so the model does not read a hurricane evacuation as ordinary Tuesday demand. Most of the accuracy gain in a first forecasting project comes from this cleanup, not from the algorithm.

Layer 2: the calendar. Day of week, month, holiday and its weekday placement, school calendars for your feeder markets, and dated local events: on the Grand Strand, bike weeks, Carolina Country Music Fest, golf tournament weeks. Keep it as a table someone owns, with next year's dates entered before the season. Late-August demand in a family beach market is mostly a school-calendar question.

Layer 3: lead indicators. The layer most operators never build, and the one that separates a forecast from a pattern. Area lodging occupancy already on the books, reservation and waitlist pace, advance ticket sales, tee times booked, and for home services the prior week's call volume. Watch the booking window itself: when it compresses, the same eventual demand reads as a collapse in pace that never arrives.

Layer 4: the near-term adjustment. Weather, last-minute events, a competitor's closure, applied in the final three to five days by a manager rather than automatically. Rain shifts demand between dayparts and between indoor and patio seating in ways a weather variable handles poorly.

Each layer needs an owner and a refresh cadence. Process first, technology second: a forecast maintained by nobody decays into a seasonal average with extra steps.

What Better Forecasting Is Worth: Worked Math for a 90-Seat Restaurant

Forecasting usually pays for itself, though the return comes from the overstaffing side rather than the dramatic understaffing stories. Consider a hypothetical 90-seat Grand Strand restaurant doing $3.2 million a year, with labor near the 36.5% median; the assumptions below are illustrative, so substitute your own. Suppose daily volume forecasts are off by an average of 18% in absolute terms and managers hedge upward, as they sensibly do when being short is visible and being heavy is not.

Overstaffing. Say 45% of days, about 164 a year, come in below forecast, each carrying roughly 1.5 unnecessary six-hour shifts: 164 × 1.5 × 6 = 1,476 excess hours. At $18 fully loaded, $26,568.

Understaffing. Say 20% of days, about 73, run far enough above forecast that the wait turns away eight covers at a $34 average check: $272 a day, $19,856 a year in forgone revenue.

Total exposure is roughly $46,400. Cutting average error from 18% to 10% does not erase it; assume you recover 40%, about $18,560 a year. Against that, scheduling software at $5 per employee per month across 40 average employees costs $2,400 a year, though the questions to ask before signing matter more than the price, and the one-time work of cleaning the data, building the event calendar, writing productivity standards, and training managers runs $6,000 to $15,000. Call it $10,000. Net recurring benefit is about $16,160, so the build pays back in roughly seven and a half months.

That figure assumes managers actually schedule to the forecast, which is a change-management problem rather than a software one, and it counts nothing for the turnover erratic schedules cause.

Where AI Helps and What It Must Not Decide

AI belongs in the prediction, not in the decision about a person's week. Machine learning weighs twenty overlapping signals consistently at three in the afternoon, which people do badly; people know what a schedule means to the person on it, which models do badly.

A 2026 study in the International Journal of Hospitality Management, "Predicting and planning: An intelligent system for demand-based hotel staff scheduling," pairs a neural network for demand prediction with optimization algorithms for scheduling and reports schedules that beat manual methods with fewer later adjustments. Note what was automated: the prediction and a proposed schedule, with people still running the floor.

Where AI earns its place: learning the interaction between weather, day of week, and event calendar that a spreadsheet cannot express; producing the confidence range instead of a point estimate; flagging days where its own prediction is unusually uncertain; and reconciling forecast against booking pace daily rather than weekly.

What stays with a person: who gets cut, who gets the extra shift, and how far ahead the schedule is published. Those carry judgment, fairness, and legal edges. Oregon's predictive scheduling law, a model other jurisdictions have followed, requires covered retail, hospitality, and food service employers with 500 or more employees worldwide to post written schedules at least 14 calendar days ahead and to pay a penalty for changing them. South Carolina has no such statute, but a system that optimizes by reshuffling shifts three days out is unusable the day you open in a jurisdiction that does, and it costs you people in the meantime. AI enhances the judgment of the manager building the plan; it does not build the plan.

Four Failure Modes and the Early Signals

Most forecasting projects fail quietly, in ways that look like success on the monthly report. Each of these has an early signal you can watch for in the first eight weeks.

1. Forecasting dollars instead of work. Signal: revenue accuracy improves but Saturday is still a mess. Fix by forecasting covers, check-ins, or tickets and converting afterward.
2. Training on a distorted history. Signal: the model misses the same weeks every year by a similar amount. Fix by relabeling closures, remodels, and one-off events in the source data rather than adding variables.
3. Averaging across dayparts. Signal: daily labor percentage looks fine while the 6 to 8 p.m. wait runs forty minutes. Fix by forecasting at the daypart the schedule is built in.
4. Unmeasured manager overrides. Signal: nobody can say whether overrides help. Log every override with a reason and review monthly. Ones that consistently beat the model are knowledge to put into it; ones that lose are a training conversation.

The fourth decides whether the project survives. A forecast managers quietly ignore is worse than none, which is why any data-run business watches that gap.

A 90-Day Build Before the Next Season

Do not build a forecast in peak season. The work needs the quiet months, and the payoff arrives next spring.

Days 1 to 30, establish the baseline. Export and clean two to three years of daily transaction data by daypart, write down the productivity standards already in use, then measure current accuracy against last season's actuals. That average absolute error is your benchmark.

Days 31 to 60, build and backtest. Add the calendar and lead-indicator layers and test against a season the model has never seen. Report accuracy by daypart and check for systematic bias, which hurts more than random error.

Days 61 to 90, run it in parallel. Produce a forecast schedule alongside the manager's for six to eight weeks without acting on it, log every disagreement, then publish. Track average absolute error by daypart, labor hours against the forecast requirement, and override frequency. Those three say within two months whether it worked, the same discipline that makes pricing decisions defensible for lodging operators in this market.

Frequently Asked Questions

Ready to Forecast Next Season Instead of Reacting to It?

Seasonal staffing is a data problem before it is a software problem: clean the history, name the drivers, then automate what remains. ANtreliX is a strategy-first AI and business intelligence consultancy in North Myrtle Beach, South Carolina, working with hospitality and tourism operators locally and remotely to build forecasts their managers will actually schedule against. See our case studies, or Talk to ANtreliX.