Forecast accuracy: measure it or stop forecasting

The manager writes a forecast every Monday: covers for each service, carefully worked out. The kitchen nods, then preps the way it always has, because it has never been told whether the forecasts were right. Nobody checks them afterwards. After a few months, the forecast is a ritual that changes nothing. Forecast accuracy is what turns a forecast from a guess on paper into a number people trust and use.

Checking a forecast is simple: compare what you expected with what happened, service by service, and look at the pattern. It takes ten minutes a week. Without it, nobody knows whether the forecast is good enough to prep from, whether it is always a little high or always a little low, or which services it gets wrong. With it, the forecast improves every week, and the kitchen can start to rely on it.

This guide covers how to score a forecast against what happened, the measures that matter, how to spot bias, the error rate at which the kitchen can trust a forecast and what to do when the forecast keeps missing. Every venue and figure in it is an example.

A forecast is only as useful as it is accurate. If the kitchen preps for 100 covers and 70 arrive, food is wasted. If it preps for 70 and 100 arrive, dishes run out and tickets slow down. Staffing works the same way: too many people on a quiet night costs money; too few on a busy one costs service.

Knowing how accurate the forecast is lets everyone decide how much to rely on it. A forecast that is usually within a few covers can drive prep directly. One that is often far off needs a margin of safety, or more work before anyone trusts it.

A forecast that is never checked is not really a forecast; it is a guess that nobody learns from. It may be consistently wrong in the same direction, and nobody will notice. It may be good, and nobody will believe it. Either way, the effort of making it is largely wasted.

So the rule is simple: if you forecast, measure the forecast. If you are not willing to spend ten minutes a week checking it, it may be better to stop forecasting and use a simpler rule, such as prepping from the same night last week, than to keep a forecast nobody trusts.

Start by recording two numbers for every service: the forecast, written down in advance, and the actual, what really happened. Covers are the simplest to compare; sales can be added later. Keep them side by side in a sheet, one line per service.

Write the forecast down before the week begins, and do not change it afterwards. A forecast that is quietly adjusted during the week, as bookings arrive, cannot be scored fairly. If you update it, keep the original too, so you can see how much each version was worth.

The error for a service is the difference between the actual and the forecast. If Friday dinner was forecast at 100 and 92 came, the error is minus 8: the forecast was 8 too high. If Saturday lunch was forecast at 60 and 66 came, the error is plus 6.

Recording the sign, plus or minus, matters, because it shows the direction of each miss. A forecast that is always too high has a different problem from one that misses in both directions.

A miss of 8 covers means different things on a quiet Tuesday lunch and a busy Saturday dinner. Percentage error puts them on the same scale: the size of the miss divided by the actual covers. Eight covers on 92 is about nine per cent; eight covers on 20 is forty per cent.

Use the size of the miss, ignoring the sign, for the percentage. That shows how far off each forecast was, whichever direction it missed in, and lets you compare services of very different sizes.

The average of those percentage errors over a week, or a month, is a single measure of how accurate the forecast is. Statisticians call it the mean absolute percentage error. In plain terms, it answers: on average, how far off was the forecast, as a share of the actual?

A forecast with an average error of ten per cent is usually within about a tenth of what happened. Track the average week by week. If it is falling, the forecast is improving. If it is rising, something has changed that the forecast is not capturing.

Bias is the tendency to miss in one direction. Add up the errors with their signs, plus and minus, over a few weeks. If the total is close to zero, the forecast misses both ways and has no bias. If it is strongly negative, the forecast is usually too high; strongly positive, usually too low.

Bias is often the easiest problem to fix. A forecast that is consistently ten per cent too high on weekday lunches can simply be lowered by ten per cent for those services, and it becomes far more accurate overnight.

Look at accuracy service by service, not just overall. A forecast may be excellent for weekend dinners, which are mostly booked in advance, and poor for weekday lunches, which depend on walk-ins. An overall average hides those differences.

Knowing which services are hard to forecast tells you where to focus. It may be worth tracking walk-ins more carefully at lunch, or noting local events that affect particular nights, rather than trying to improve services that are already accurate.

Forecast variance, how widely the misses spread, matters as much as the average. A forecast that is usually within five covers but occasionally misses by forty is harder to rely on than one that is always within ten. The occasional big miss is often what causes the worst nights.

Look at the biggest misses each month, separately from the average. Each one usually has a reason, such as an event, a large party or a closure nearby, and those reasons are the best clues for improving the forecast.

How accurate does a forecast need to be before the kitchen can prep from it? There is no universal answer, but a useful rule of thumb is that a forecast usually within about ten per cent of what happens is close enough to prep from, with a sensible buffer for dishes that cannot be made quickly to order.

A forecast that is often twenty per cent or more off needs a bigger buffer, or more work, before the kitchen can lean on it. Share the accuracy with the kitchen, so the team knows how much to trust each service's forecast rather than having to guess.

Prep from forecast means deciding how much to prepare from the forecast covers and the usual share of each dish, rather than from yesterday's prep or a feeling. It is where forecast accuracy pays for itself, through less waste on quiet nights and fewer shortages on busy ones.

The more accurate the forecast, the closer prep can follow it. Our guide to prep from forecast, not from yesterday's prep covers how to turn forecast covers into a prep list.

A buffer is extra prep above the forecast, to cover the chance that more guests come than expected. Size it to the forecast's typical error. If a service is usually within ten per cent, a ten per cent buffer covers most nights. If it is usually within twenty, a larger buffer is needed.

Buffers can also vary by dish. Items that are quick to make to order need little buffer; slow items, such as braises, stocks and anything that takes hours, need more, because they cannot be made in a hurry if the night turns busy.

A covers forecast is the foundation for everything else: prep, staffing and sales. Its accuracy deserves the most attention. Check it service by service, record the misses and look for the patterns behind them.

When the covers forecast is accurate, the rest follows more easily. Our guide to a covers forecast your kitchen will believe covers how to build one the kitchen trusts.

Sales forecasting adds a second source of error: spend per cover. Even if covers are forecast exactly, sales can miss if guests spend more or less than expected, for example on a night with a set menu or a large party on a fixed package.

Check sales accuracy separately from covers. If covers are accurate but sales are not, the problem is the spend per cover assumption, not the guest count. Adjust spend by service and by kind of night, and the sales forecast improves.

Keep a short weekly scorecard, one line per service:

At the bottom, the week's average percentage error and the total of the signed errors, to show bias. It takes ten minutes after the week ends, and over a month it shows exactly how good the forecast is.

The notes beside the big misses are the most useful part of the scorecard. "Concert at the arena, not in the calendar." "Large party booked on the day." "Heavy rain, terrace closed." "Competitor opened round the corner." Each explains a miss and suggests how to avoid the next one.

Turn the lessons into changes: add the arena's calendar to the events list, check for large late bookings before finalising Friday's prep, add a rain adjustment for terrace services. Over time, the same kinds of miss stop happening.

Big misses usually come from a handful of causes: local events nobody noted, large parties booked late, holidays and school breaks, unusual weather for venues that depend on it, a nearby closure or opening, and one-off disruptions such as roadworks or a transport strike.

Keep a list of the causes you have found, and check it each week when you forecast. Many big misses are predictable once you know to look for them.

To improve accuracy, look at which input caused the misses. If bookings already made were right but pick-up was underestimated, adjust the pick-up. If walk-ins were overestimated at lunch, lower them. If a season turned earlier than last year's pattern suggested, give recent weeks more weight.

Change one thing at a time and watch the scorecard. If the average error falls, keep the change. If it does not, try something else. Our guide to demand forecasting without a data scientist covers the inputs that matter most.

There is a point where further effort is not worth it. A forecast that is within a few covers on most services, with no strong bias, is doing its job. Spending hours each week to shave another point off the error rarely pays.

Once the forecast is good enough for prep and staffing, keep the routine going, keep checking and only investigate when the error rises or a big miss appears. The aim is a useful forecast, not a perfect one.

If, after several months of checking and adjusting, a forecast is still no better than a simple rule, such as "the same night last week", it may be time to stop forecasting and use the simple rule instead. Measure the simple rule the same way: if it is as accurate, it is cheaper.

This rarely happens for venues with regular trade and a steady book of reservations, where a forecast usually beats simple rules. It is more likely for venues whose trade is almost entirely walk-ins and unpredictable events.

A new venue's forecasts will be less accurate at first, because there is little history to work from. That is normal. Check them every day rather than every week in the first month, and adjust quickly as patterns appear.

Use bigger buffers until the forecast settles, and share the uncertainty with the kitchen. Within a few months, the venue's own patterns usually make the forecast far more reliable.

Staffing decisions depend on the forecast too, and the cost of a miss is different. Too many staff on a quiet night costs wages; too few on a busy night costs service and puts pressure on the team. Knowing the forecast's accuracy helps decide how much flexibility to build in.

For services that are hard to forecast, consider staff who can be called in or sent home early, within fair working practices. Our guide to scheduling from forecast, not from memory covers how to build a rota from the forecast.

The kitchen will trust the forecast more if it can see how accurate it has been. Share the weekly scorecard with the head chef, or at least the average error and the big misses with their reasons. A chef who sees that Saturday forecasts have been within five covers for two months will prep from them with confidence.

Ask the kitchen for its view too. Chefs often know things the manager does not, such as a large party mentioned at the pass or a regular group that always comes on the first Friday of the month.

Track the average error over months, not just weeks. A forecast that was ten per cent out in spring and six per cent out in summer is improving. One that was six per cent out and is now twelve has stopped capturing something: a change in the venue, the neighbourhood or the guests.

A simple chart of the monthly average error, kept with the scorecards, shows the trend at a glance. It is one of the best measures of whether the forecasting routine is working.

Services that depend on walk-ins, such as a city-centre lunch, are usually harder to forecast than booked services, because there is less to go on in advance. Expect a higher error for them, and do not judge the whole forecast by them.

Track walk-ins carefully for these services, by day and by weather, for a few months. Patterns often appear, such as walk-ins rising on sunny days or falling during school holidays, and adding them can bring the error down.

Special nights, such as a festive party evening or a one-off event, are hard to forecast because there is little history. Treat their forecasts as estimates, score them separately from ordinary services and do not let them distort the average.

Keep a note of how each special night went compared with the forecast. When the same kind of night comes round again, the note is the best guide you have.

An invented bistro has forecast covers for months without checking. The manager starts a weekly scorecard. After four weeks, the average percentage error is about fourteen per cent, and the signed errors show that weekday lunches are consistently forecast around twelve per cent too high, while Friday dinners are often too low.

She lowers the weekday lunch forecasts and adds Friday's usual late pick-up. The next month, the average error falls to about eight per cent, with no strong bias. The head chef starts prepping from the forecast with a ten per cent buffer on slow dishes, and the kitchen's end-of-night waste falls noticeably.

servearound does not store forecasts or score them, so the forecast and its error are kept in your own sheet. What it gives you is the actual side of the sum: reservation analytics show covers day by day and by meal period for any range of dates, with walk-ins as their own source, and payment analytics show what was collected by day, so last week's covers and sales for each service can be set against what you forecast.

Kitchen analytics show tickets completed and which items cost the most time, and the menu report shows units sold for every dish over any dates, so the prep side, how many of each dish the covers turned into, can be checked too. See Analytics.

Because a forecast nobody checks cannot be trusted, improved or used with confidence.

Record the forecast and the actual for each service, and calculate the error and the percentage error.

The average size of the misses as a share of the actual, over a week or a month.

A tendency to miss in one direction, shown by adding the errors with their signs.

A rule of thumb is usually within about ten per cent for the kitchen to prep from it with a sensible buffer.

Write down the reason, and add it to the inputs or the events calendar for next time.

If, after months of checking, a simple rule is just as accurate.

Yes. Trust in the forecast comes from seeing how accurate it has been.

Forecast accuracy is what makes a forecast worth making: measure it or stop forecasting. Write the forecast down in advance, record what actually happened for each service, and calculate the error, the percentage error and the average over the week. Add up the signed errors to spot bias, and look at accuracy service by service and at the biggest misses with their reasons.

A forecast usually within about ten per cent is close enough for the kitchen to prep from, with buffers sized to each service's typical error. Fix bias first, learn from every big miss, change one input at a time and share the scorecard with the kitchen. Track the average error over months, and if a simple rule turns out just as accurate, use the simple rule.