How to Forecast Production and Packaging Quantities for Multi-Site
How to Forecast Production and Packaging Quantities for Multi-Site Catering, With Confidence Ranges
Multi-site caterers do not have a demand problem and a packaging problem; they have one forecasting problem that ends in a pallet. When you hold historical covers, reservations, weather feeds, class schedules, and meeting calendars, the disciplined path is to define one forecast target per site, meal period, and date, build simple baselines first, add regression on external predictors where they earn their keep, publish prediction intervals rather than single numbers, reconcile those intervals
Define the Forecast Target Before Anything Else
One row in your data should mean one thing:
service date × site × meal period → actual covers
For a contract caterer running a hospital site, a campus site, and two corporate offices, a single Monday produces six to twelve of these rows — breakfast, lunch, dinner per site. Everything downstream depends on keeping this grain clean. If one site logs catering trays separately from retail counter service, split them into separate meal periods rather than blending them; a blended average destroys the reservation and calendar signals you are about to use.
Write the target down as a table and treat missing rows as zero-service days, not gaps to be interpolated. A site that closed for a holiday served zero covers, and your model needs to see that.
Start With Simple Baselines, Then Add Predictors That Earn Their Keep
Baseline: same weekday, trailing average
For each site × meal period cell, the first model is boring on purpose: the mean or median of the same weekday over the last four and eight weeks, adjusted for obvious holiday effects. This baseline is your sanity check. If a complex model cannot beat it on held-out weeks, the complexity is noise.
Regression with external predictors
Your second layer is a regression per cell (or a pooled model with site and period as categorical features) using four predictor groups:
Historical demand. Covers at t−7, t−14, t−21; same-weekday averages; recent trend; variance around the recent level.
Reservations. Current bookings, reservation pace compared with the same lead time in past weeks, party counts, cancellation and no-show rates, large events. Reservation pace is usually more informative than the raw count — how fast bookings are accumulating tells you more than how many exist today.
Calendar. Weekday, holidays, academic term/exam/break status, scheduled meetings and events with expected attendance, and whether an event overlaps breakfast, lunch, or dinner. For campus sites, enrollment by time block matters.
Weather. Temperature, precipitation probability, snow, and severe-weather flags versus seasonal norms. Weather forecasts are legitimate forward-looking predictors, but remember the weather forecast itself is uncertain — widen your intervals a little on days the forecast is far from the seasonal norm or issued with low confidence.
In a catering context these predictors are unusually good because most are known before production: reservations and calendars especially. That is why site × meal-period forecasts for food service can be materially tighter than generic retail demand forecasts.
Publish Prediction Intervals, Not Single Numbers
A point forecast invites one procurement decision and one production batch, which is exactly wrong when demand varies. Produce P10/P50/P90 per cell:
Read it as: in nine weeks out of ten, Monday lunch lands between 1,050 and 1,330 covers. The width of that band is itself information — a hospital breakfast band of ±4% supports just-in-time replenishment; a corporate lunch band of ±25% around a meeting-heavy week says buy flexibility, not precision.
Standard practice from statistical forecasting is to use prediction intervals rather than bare point estimates precisely because they carry the uncertainty explicitly. One caution: intervals built naively from regression residuals tend to be too narrow, because they ignore the uncertainty inside the predictor inputs (the weather forecast) and the possibility the model itself is misspecified. Inflate conservatively and validate the coverage empirically — check what share of past weeks actually fell inside the P10–P90 band. If fewer than ~80% did, your intervals are optimistic.
Reconcile Up the Hierarchy
Independent cell forecasts will not add up: sum the six Monday-lunch cells and the total may contradict your site-level and company-level expectations. Reconciliation fixes this, and it matters operationally because you buy packaging at the site level and negotiate at the company level.
A practical approach:
Forecast every cell independently.
Compute aggregates (site-day, site-week, company-week) from those cells.
Apply a reconciliation step — proportional adjustment of cells toward the more reliable level, or a formal hierarchical method — so bottom and top agree.
Apply reconciliation to the whole distribution, not just the medians. If you reconcile only P50s, your P90 at the company level will be understated and your safety stock will be quietly wrong.
Version the Forecast: D-7, D-3, D-1
Run the forecast at fixed horizons and keep every version:
D-7 (a week ahead): the planning forecast. Drives raw-material orders and packaging replenishment with lead time.
D-3: the commitment forecast. Drives production schedules and any late packaging orders.
D-1: the execution forecast. Drives batch sizes, staff rostering, and final pull from packaging safety stock.
Versioning is not bureaucracy — it is your measurement system. By comparing D-7 against D-3 against D-1 against actuals per cell, you learn how quickly each site's demand resolves, and that tells you where forecast confidence genuinely improves with lead time and where it never does. Sites whose D-7 error barely exceeds their D-1 error are calendar-driven and should be planned on the weekly cycle; sites where error collapses between D-3 and D-1 are reservation-driven and reward a late packaging top-up lane.
Track Forecast Error Against Waste
Error metrics (MAPE or weighted absolute error per cell) belong on the same dashboard as food waste and packaging usage. The connection is direct: over-forecast → surplus prepared covers → wasted food and wasted bowls and lids; under-forecast → stockouts, substitutions, and emergency purchases at bad prices.
Track two ratios per site and period:
Forecast bias (systematic over or under). Persistent bias means a predictor is missing — usually a calendar event type.
Waste per forecasted cover. Falling waste at flat accuracy means your intervals are being used well in production decisions.
Without the packaging dimension, forecasting projects stall at the kitchen door. With it, you can show procurement that a tighter lunch band at one campus reduced both short-life waste and the cartons of unused 1000ml kraft bowls opened that month.
Translate Covers Into Packaging Quantities Per Forecast Band
This is where forecasting becomes a purchasing system. Each meal-period cell maps to a packaging bill of materials: covers → units of bowl or container → matching lids → bags and sleeves where the service model needs them. A 1,000-cover lunch with 70% hot-bar share in 1000ml kraft paper bowls is roughly 700 bowls plus 700 lids (lids are typically ordered separately); the cold line may draw on clear PET cups; the executive meeting overlap may pull japanese-style single-compartment containers for plated service.
Then let the forecast band drive the order quantity:
P50 covers × BOM = base order quantity.
(P90 − P50) covers × BOM = the quantity held as packaging safety stock or covered by a scheduled top-up between D-7 and D-3.
Packaging has one advantage food does not: it does not spoil. So the correct posture is usually to carry the band, not gamble on the point estimate. The cost of holding an extra 15% of bowls and lids for a week is small next to a stockout at a 1,300-cover lunch. Size the safety stock by the observed interval width per cell — wide-band cells carry more, tight-band cells carry less — and set reorder points from the D-7 forecast plus that band rather than from flat historical averages.
Two operational notes from the supplier side:
Order per forecast band, not per hunch. Handing your packaging supplier the P10/P50/P90 table by site and period lets them plan production slots and hold or release stock against your actual draw pattern. TakeawayPack works with contract caterers and central kitchens on exactly this: quantities planned per meal period, with kraft paper bowls (500–1500ml), kraft lunch boxes, and PP container lines ordered against forecast bands and replenished on reorder points rather than ad-hoc calls.
Consolidate SKUs where the menu allows. Fewer bowl sizes and lid types across sites means each forecast band aggregates into deeper, more predictable orders — better fill rates for you and more efficient production runs for the factory. Mixed-SKU cartons (bowls plus their lids, or a site-week assortment packed together) reduce the split-carton handling that otherwise multiplies when six sites order small quantities independently.
A Weekly Operating Rhythm
Pull it together into one cycle:
Monday, D-7: reconcile next week's cell forecasts; export P50 and P90 tables; release packaging replenishment orders against the bands.
Thursday, D-3: refresh with reservation pace; adjust production schedule; place top-up packaging orders for cells whose P90 moved up.
Daily, D-1: final batch sizes; pull from packaging safety stock; log actual covers, food waste, and packaging usage.
Weekly review: compare forecast versions against actuals per cell; check interval coverage; adjust safety-stock multipliers where bands are systematically too wide or too narrow.
The output is a procurement and production system that speaks in probabilities end to end — from next Monday's campus lunch to the carton count on the pallet.
Plan Your Packaging Around the Forecast
If you are building this forecast for multiple sites and want packaging supply that follows the same bands instead of fighting them, Talk with TakeawayPack at takeawaypack.com. We supply kraft paper bowls with separately ordered lids, kraft lunch boxes, american-style rectangular PP containers, japanese-style single-compartment containers, clear PET/PP cups, and matching bags and sleeves, and we can plan mixed-SKU cartons and replenishment against your forecast ranges per site and meal period.
