Almost every restaurant forecasts sales. Very few forecast them usefully. The common version is a number written on the top of the rota on a Monday, arrived at by glancing at the same week last year, adding a bit for optimism, and never once going back to see whether it was right. That is not a forecast, it is a wish with a decimal point. A real forecast is a testable claim about how many people will walk in and what they will spend, built from the sales history sitting in your point of sale system, and judged afterwards on how close it came.
The difference matters because the forecast is not the deliverable. The rota is, and the order is. A forecast exists to decide two things: how many labour hours to schedule, and how much food to buy. Get the projection roughly right and both of those decisions get quietly easier, which is why forecasting sits upstream of stock control rather than beside it. Get it wrong in the same direction every week and you are either paying staff to stand still or sending guests away, and you will not know which until the month closes. This is the weekly method: covers first, then spend, then the honest arithmetic of how wrong you were.
Why most restaurant forecasts fail
There are three failure modes and they are all procedural rather than mathematical. The first is forecasting revenue directly. Revenue is a product of two things that move independently, so a single revenue number hides which one moved. If you projected a certain figure and missed it, you cannot tell from the miss alone whether fewer people came or the people who came spent less, and those two problems have entirely different fixes.
The second is treating the forecast as a target. A target is what you would like to happen and it belongs in the budget. A forecast is what you expect to happen given what you know today, and the two should be allowed to disagree. When managers are appraised on hitting the forecast, the forecast drifts upward to look ambitious, then the rota is built on it, and you have overstaffed on the strength of a morale exercise. Keep the budget conversation and the forecasting conversation in separate meetings if you can.
The third and most common is never scoring it. Nobody goes back on Tuesday and asks how far off last Wednesday was. Without that loop the forecast never improves, because the only mechanism by which a forecast gets better is somebody noticing a pattern in its mistakes. A forecast you do not measure is a forecast you cannot trust, and one you cannot trust will eventually be overridden by whoever is most confident in the room.
Forecast covers first, then average check
Split the number in two and forecast each half separately. Covers are a demand signal: how many guests you expect through the door, by service. Average check is a monetisation signal: what each of them is worth once they sit down. Multiply them and you have projected sales, but you have also kept the two levers visible, which is the whole point.
Covers are the harder and more valuable half, because covers drive almost every operational decision downstream. Labour hours track covers, not revenue. Prep volumes track covers. Table availability, waitlist pressure and kitchen ticket times all track covers. A forecast expressed in covers can be handed straight to a head chef and a floor manager and acted on without translation, whereas a revenue figure has to be converted by each of them before it means anything.
Average check is more stable than most operators expect, which is good news, because it means you can usually carry last period's figure forward with a small adjustment and be close. It moves for identifiable reasons: a menu price change, a shift in mix towards or away from the higher margin dishes, a change in how hard the team is selling, or a change in who is coming in. Track it as its own metric, as covered in average check and how to raise it, and you will notice those shifts as they happen rather than in the quarterly accounts.
Building the baseline from your own sales history
Pull at least twelve months of daily sales if you have it, and two or three years if the business has been open that long and has not changed format. Twelve months is the minimum that lets you see a full seasonal cycle; anything shorter and you will mistake a seasonal trough for a decline. Export it by day rather than by week or month, because the day is the unit you actually schedule and order against.
Shape the export so that every row carries the date, the day of the week, covers, net sales, and ideally the daypart and the channel. The day of the week is the single most predictive column you have. Restaurant demand is far more weekly than it is monthly: a Tuesday resembles other Tuesdays much more than it resembles the Monday next to it, so the useful comparison is always same day to same day. This is why a monthly average is nearly useless for scheduling and why the four-week rolling average became the standard starting point.
The rolling average is simple enough to do in a spreadsheet. For next Wednesday, take the covers from the last four Wednesdays and average them. It is crude, it lags a genuine trend by a couple of weeks, and it is still better than instinct for most single-site restaurants. If you want the arithmetic laid out with the labour and cost ratios attached, the operator calculators do the same job without the spreadsheet.
Stripping the noise before you average anything
An average is only as good as the days going into it, and restaurant weeks are full of days that should not be in there. A private buyout, a public holiday, a snow day, a power cut, a road closure, a one-off promotion, the week the dining room was half shut for repairs: every one of these will drag the average towards a reality that is not coming back next week.
Take a worked example. The last four Wednesdays produced 96, 104, 88 and 172 covers. Averaged naively that gives 115, and you would staff and order for 115. But the 172 was a private party that filled the back room, which you know because somebody wrote it down. Strip it out and the average of the remaining three is 96, which is very close to a normal Wednesday. The naive figure would have had you nearly twenty per cent overstaffed and holding stock you did not need.
This only works if the exceptions get recorded at the time, which makes the log the actual deliverable here. Keep a running note against each unusual day saying what happened, and keep it somewhere the next manager will find it. Most teams already have a natural home for this in the shift handover notes or the closing checklist. A year of those notes is worth more to a forecast than any amount of statistical technique, because nothing recovers the reason for a strange Tuesday once everybody has forgotten it.
Layering in what you already know is coming
The baseline tells you what a normal week looks like. Next week is not a normal week, and you usually know several of the reasons why before it starts. This is the part where a human operator still comfortably beats a naive model, because much of the relevant information is local and never appears in the sales history.
Work through a short standing list. Public holidays and the days either side of them, which often move demand rather than adding it. School terms and half terms, which reshape family trade and lunch entirely. Local events with a known footfall: a match, a concert, a conference, a festival, a market day. Weather, which matters enormously if you have outside covers and barely at all if you do not. Roadworks and transport disruption. Whatever your neighbours are doing, since a new opening two doors down will take a slice of your trade for a few weeks. And your own actions: a menu change, a price change, a campaign, a closure.
Quantify each one as a percentage adjustment rather than a new absolute number, because percentages are easier to check afterwards and easier to reuse. Carry on the example: the baseline Wednesday is 96 covers, and a conference in town has historically added around ten per cent to midweek trade, so the forecast becomes 106 covers. Write down both the 96 and the ten per cent separately. When the actual comes in you can then see whether the baseline was wrong or the conference assumption was wrong, and only one of those is worth arguing about.
Reservations are a partial read on all of this and should be used carefully. A booked cover is much better evidence than a guess, but the ratio of bookings to walk-ins varies by day and by season, so the booking count on Monday for the coming Saturday is only meaningful against how that count normally converts. If you take reservations at volume, your reservation system already holds the pace-of-booking curve you need, and the waitlist tells you which services turned people away, which is demand your sales history does not contain.
Dayparts and channels: where a daily number hides the problem
A single daily figure is fine for a site that does one service. For anything more complicated it averages away the thing you needed to see. If lunch is soft and dinner is strong, the day looks acceptable and you will keep staffing lunch as though it were fine. Split the forecast by daypart wherever the dayparts have different staffing patterns, which in practice means almost always.
Split by channel too, if you have more than one. Dine-in, takeaway, delivery and any counter or kiosk trade behave differently and have very different labour and packaging consequences per unit of revenue. Delivery in particular tends to be countercyclical to dine-in on bad weather days, so a combined number can look stable while the mix underneath it swings hard, and the mix is what decides whether you needed another kitchen hand or another server. Where a meaningful share of orders arrives through third-party delivery platforms, forecast that stream separately, because its average check and its margin are both different.
The counterargument to all this splitting is that thin data gets thinner every time you divide it. That is real. If a daypart runs twelve covers, forecasting it separately is false precision. The rule of thumb worth applying is to split only where the split changes a decision: if lunch and dinner are scheduled and prepped separately, forecast them separately, and otherwise leave them together.
Measuring how wrong you were
This is the step nearly everyone skips, and it is the step that turns forecasting from an opinion into a skill. Every week, put the forecast and the actual side by side and record the gap. Do it as a percentage of the actual so the numbers are comparable across a quiet Monday and a heaving Saturday.
Carry the example forward. You forecast 106 covers for Wednesday and 121 walked in. The gap is 15 covers, which against an actual of 121 is a miss of a little over twelve per cent. Do that for all seven days, take the average of the misses while ignoring whether each was high or low, and you have a single number describing how accurate your forecasting currently is. A week of misses of six, eleven, four, twelve, nine, three and seven per cent averages out at around seven and a half per cent.
What counts as good depends on your format and volume, and anyone quoting a universal figure is selling something. As a working guide, a single site with steady trade that consistently lands inside ten per cent on covers is forecasting well enough to schedule confidently, and inside five per cent is genuinely good. High-volume sites with stable patterns do better; small sites, weather-exposed sites and anywhere with lumpy group business do worse, and no technique will fix that entirely.
Then look separately at direction, because accuracy and bias are different faults. Add up the misses keeping their signs this time. If they roughly cancel out, your forecast is noisy but centred, and the fix is better information. If they all lean the same way, you have a bias, and the fix is arithmetic: you are systematically under or over and can simply correct for it. Persistent underforecasting is the more expensive of the two, because it shows up as understaffing, slower service and lost covers rather than as an obvious cost line. Track the pair of them alongside your other operating KPIs and the improvement becomes visible within a couple of months.
Turning the forecast into a rota
A forecast that does not change the schedule has not done anything. The link between the two is a productivity target, and there are two common ways to express it. The first is labour as a percentage of sales, which is the language most owners and accountants already use. The second is a volume measure such as sales per labour hour or covers per labour hour, which is more useful to a manager because it does not move when you change menu prices.
Work it through with the numbers already in play. The forecast is 106 covers at an average check of 31.50, giving projected sales of 3,339. If your labour target for that day is 28 per cent of sales, that allows 934 in wages, and at a blended rate of 14 an hour that is roughly 67 hours to spread across the day. Whether 67 hours is the right shape is a separate question from whether it is the right quantity, and the shape is where most of the money is won or lost: the same total hours arranged badly leaves you three deep at four in the afternoon and drowning at eight.
Which is why the daypart split earns its keep here. Distribute the hours against the forecast curve rather than evenly, then sanity check the result against the constraints a spreadsheet does not know about: minimum shift lengths, who is trained on what station, availability, and the legal rest requirements in your jurisdiction. The mechanics of building the rota itself, including how to handle the constraints, are covered in staff scheduling, and how to read the resulting cost line is in restaurant labour cost.
One caution on labour percentage as a target. It flatters you on a strong day and punishes you on a weak one, because a large slice of your labour is fixed regardless of how many people come in. A manager who is held to a flat percentage on a quiet Tuesday will cut the shift below what the service actually needs. Set the target by day or by daypart rather than as a single site-wide figure, and it stops fighting you.
Turning the forecast into a purchase order
The same projection drives buying, and here the conversion runs through the menu. If the forecast is 106 covers and a given dish historically sells to eighteen per cent of covers, you need about 19 portions of it. Do that across the movers, add whatever is already in the walk-in, and the order more or less writes itself.
Two adjustments make this survive contact with reality. First, hold the safety margin only on the fast movers and the items with a long lead time, not across the board, since a blanket buffer on everything is just a decision to carry more stock than you need. Second, weight the margin by shelf life. Being short of a dry good is an irritation and being short of the dish everyone came for is a lost sale, but over-ordering fresh fish is money in the bin, so the buffer should be generous on the stable items and tight on the perishable ones. That tradeoff is the practical heart of inventory management and the main lever on reducing food waste.
Over time the forecast should set your par levels rather than the other way round. Par levels drawn up once and never revisited slowly detach from the business as the menu and the trade change, and the symptom is a stockroom holding plenty of what no longer sells. Reviewing pars against the rolling forecast a few times a year, alongside the supplier terms and order schedules described in procurement and vendor management, keeps the two in step.
Forecasting a site with no sales history
A new opening has no history to average, so the first forecast is built from structure rather than from data. Start with physical capacity, which is knowable: the number of seats, the number of services, and a realistic assumption about how many times each table turns. Seats multiplied by turns gives a theoretical ceiling for the service, and table turnover covers how to arrive at a turn assumption that is not simply hopeful.
Then apply a utilisation rate, because nobody fills the room every service, and be pessimistic about the opening weeks in both directions. Many new restaurants see an initial curiosity spike that decays over the first month or two, followed by a slower climb as word spreads, so the first four weeks are the worst possible basis for a permanent forecast. Assume the early numbers are unreliable and revisit them monthly.
For average check, price your own menu against the mix you expect rather than borrowing an industry figure, since check is the number most sensitive to your specific pricing and format. Model it low, mid and high instead of committing to a single value, and carry all three into the cash plan. The wider financial framing for this sits in the restaurant business plan and startup costs, and the number that decides whether the model survives is in break-even analysis. Swap the structural forecast for a data-driven one as soon as you have eight to twelve weeks of trading, and do not get attached to the original.
The weekly cadence and who owns it
Forecasting fails more often on ownership than on method. It needs one named person, a fixed slot in the week, and a place the output lives where the chef and the floor manager will both see it. If it is everybody's job it is nobody's, and it will quietly stop happening in the second busy week of December.
A workable rhythm is a single session early in the week, before the rota is published and before the main order goes in, because a forecast produced after those two decisions is decoration. The session has four parts and takes well under an hour once it is a habit: score last week by putting forecast against actual and noting the misses; explain any miss over roughly ten per cent, in one line, in writing; build next week's baseline from the rolling average with the exceptions stripped; then layer the known events onto it and publish. Push the outline two or three weeks ahead in less detail so that ordering and holiday approvals have something to work against.
Across several sites the same cadence holds but the aggregation changes, and the trap is forecasting the group and dividing down. Local demand patterns differ enough between sites that a group figure split by historical share will be wrong everywhere at once. Forecast each site and add them up, which is the same principle as everything else in multi-location management: the site is the unit of operational truth and the group is a summary of it.
Seasonal venues and the longer view
Weekly forecasting handles the rota and the order. It will not tell you whether you can afford the refit, and a business with a strong season needs a second, longer horizon alongside the weekly one. A thirteen-week rolling projection is the usual compromise: long enough to see a season change coming and to plan hiring and cash around it, short enough that the assumptions are still worth something.
For a genuinely seasonal site the year-on-year comparison matters more than the rolling average, because a four-week average is actively misleading when you are climbing into or falling out of a peak. In those weeks the same week last year, adjusted for the growth you have seen in comparable recent weeks, is the better baseline. Peak trading periods deserve their own treatment, which is why festive season planning and the tactics for filling empty tables on slow nights sit either side of the forecast rather than inside it.
The longer view is also where the forecast meets the money. Projected sales feed the cash position, the wage bill, the supplier payments and the tax set-asides, and a seasonal business can be profitable on paper and still run out of cash in the shoulder months. That linkage is the subject of cash flow management, and the monthly reconciliation of forecast against outcome belongs with the profit and loss statement rather than in the weekly operational file.
What forecasting will not do for you
It will not make demand appear. A forecast is a planning instrument, and an accurate forecast of a bad week is still a bad week; the response to a soft projection belongs with increasing restaurant sales rather than with the forecast itself. Nor will it survive a change of format. Change the menu substantially, reposition the site, alter the opening hours or the price architecture, and your history stops describing your business. Expect a couple of months of poor accuracy after any real change, and say so in advance so that nobody reads it as a failure of the method.
It will not beat a hard capacity limit either. If the room seats sixty and the forecast says ninety, the forecast is telling you about demand you cannot serve, not revenue you are going to bank. That is still useful information, because suppressed demand is an argument about hours, layout, turn times or a waitlist, but it is not sales. And it will not compensate for bad data: if covers are recorded inconsistently, if voids and comps distort the sales line, or if walk-outs are never logged, the forecast inherits all of it. Reliable numbers at the point of capture, of the kind discussed in voids and comps and legible in the daily Z report, are a precondition rather than a refinement.
On software: a good spreadsheet built on clean history and a disciplined weekly habit will outperform an expensive tool used carelessly, and most single sites do not need to buy anything to start. What tooling genuinely buys you is scale and the variables you cannot be bothered to maintain by hand, chiefly weather and event feeds, which is where the automated approaches described in AI in restaurants start to pay for themselves across a group. Begin with the habit. A forecast you actually score every week will teach you more about your restaurant in a quarter than any model you do not understand.
