Demand forecasting

Know the room before it arrives.

A fortnight of covers, read shift by shift — closed Mondays, the weekend lift, lunch against dinner — turned into kilograms to prep and hands to roster. It advises; the kitchen still signs off. Read by Eleanor Marlowe, Chef & Thomas Finch, Front of house

Forecast covers · next 7 days

196

Tue–Sun; Monday closed. ~SGD 80 average spend

Forecast accuracy

~11% MAPE

Directional — mean error over the last 8 weeks

Busiest service

Sat dinner

Forecast 34 covers on 7 Jun — the fortnight's peak

Prep change vs last week

+18%

Driven by next Saturday's F1-weekend lift

The next fortnight, shift by shift

Covers per day, lunch and dinner stacked. Whiskers show the 80% confidence band (low–high). Mondays closed.

Forecast covers →

Closed

Tue3

Wed4

Thu5

Fri6

Sat7

Sun8

Mon9

Tue10

Wed11

Thu12

Fri13

Sat14

Sun15

Mon16

June 2026 →

Lunch covers

Dinner covers

Forecast peak — Sat 7 Jun (F1 weekend)

80% confidence band (low–high)

What this drives

Forecast covers converted to prep at fixed portion ratios. Next service window: Sat 7 Jun dinner, 28 covers.

Prep par — Sat 7 Jun dinner

Kilograms and portions, not guesswork

Dish	Sell-through	Portions to prep	Quantity	vs last Sat
Pie floaterBeef cheek pithivier · signature	70%	20	6.0 kg cheek	+3
Barramundi en papilloteLemon myrtle beurre blanc · signature	75%	21	9.0 kg fillet	+3
Manuka-smoked duckDavidson plum gastrique	60%	17	17 breasts	+2
Oysters Coffin BayFinger lime, champagne mignonette	55%	15	9 dozen	+2
Venison WellingtonFor two · juniper jus	35%	5	2.5 kg loin	+1
Lamington souffléRaspberry coulis, coconut sorbet	60%	17	17 ramekins	+2

It advises, it does not act. These pars are a draft for the chef to amend — they are not sent to suppliers and nothing is ordered automatically. Portion ratios are fixed against forecast covers; the kitchen has the final word on every line.

Roster suggestion — Sat 7 Jun

Hands against the peak

Kitchen — dinner service 5 on

Front of house — dinner 4 on

Suggested change vs a normal Sat +1 each

Covers per FOH head ~9

An extra chef and an extra runner cover the F1 lift without over-rostering the quieter Tuesday. A suggestion only — Thomas confirms the roster; the model never books a shift.

An anomaly the model has already accounted for

Where the forecast departs from the ordinary weekend pattern — and why.

Flagged — next Saturday

F1 weekend lifts Sat 7 Jun to 34 covers — +18% on a normal Saturday.

The baseline Saturday pattern would land near 29 covers. The model has added roughly five on the back of the Singapore Grand Prix weekend in the events calendar, two confirmed birthday bookings already in the book, and a clear-weather forecast. The lift sits in the dinner service, which is why the prep and roster changes above are concentrated there.

+3F1 weekend — Marina Bay street circuit, historically lifts CBD dinner demand
+2Two birthday bookings — parties of 6 and 8 already confirmed in the book
+0Clear weather — no rain penalty applied to walk-in estimate

Forecast covers

Confidence band

30–39

Model confidence

Medium

The number carries uncertainty. An F1-weekend lift is real but hard to size from history — it could land anywhere from 30 to 39 covers, so the band is wider than a normal Saturday. The model has flagged it; it has not assumed it.

The guardrail

Advisory only. A human signs off.

The forecast and its prep sheet are a starting point, not an instruction. Eleanor approves the prep pars and Thomas confirms the roster before either takes effect — the model proposes, it never commits.

Every forecast is logged against what actually walked in. When the model is consistently wrong — an F1 lift that didn't materialise, a rained-off Friday — the error-tracking loop surfaces the drift so the next fortnight's numbers, and the human reading them, are better calibrated.

If the room comes in light, the prep was a draft and the roster was confirmed by a person who can read the night — not a figure the kitchen was forced to commit to.

Under the hood What this is today — and where it should go next

Right now this is a concept mock-up over synthetic data: a fixed fortnight of forecast covers built from the same room logic we seed the demo with — closed Mondays, weekends heavier, dinner outweighing lunch, party sizes mostly two to six. The numbers here are illustrative, not the output of a trained model. A real build pairs a statistical forecaster with an LLM that turns the prediction into a prep sheet and a plain-language brief.

On the roadmap

A model that learns the room. Train a time-series forecaster — Prophet or gradient-boosted trees over 12+ months of POS history and the reservation book — to predict covers per shift with a genuine confidence band, rather than the hand-set band shown here.
Wire in the exogenous drivers. A live events calendar (F1, public holidays, nearby conferences) and a weather feed, so a lift like next Saturday's is detected and sized from data — not narrated after the fact.
Close the error loop. Log every forecast against actual covers, track MAPE shift by shift, and surface drift when the model is consistently wrong so it — and the chef — recalibrate.
True prep economics. Fold yield, trim and spoilage into the portion ratios so the prep sheet reflects real kilograms ordered, and feed the over/under back as an avoidable-waste figure in dollars.
The guardrails don't move. The forecast stays advisory. It can propose pars and a roster; a person still approves both, and nothing reaches a supplier or a shift automatically.

Full use-case write-up — the problem, benefits and vendor landscape: UC03 — Demand forecasting → Concept mock-up — illustrative data, not a live system.