Demand forecasting without a data scientist

Every Monday, the head chef asks how busy the week will be, and every Monday the manager gives a slightly different kind of answer: a feeling, a look at the bookings, a memory of last year. Some weeks the kitchen over-preps and throws food away; other weeks it runs out of the special by eight. Demand forecasting sounds like work for a data scientist with a computer model. For a restaurant, it is mostly four simple inputs, looked at the same way every week.

Most of the difference between a quiet week and a busy one comes from a few predictable things: which night of the week it is, how many bookings are already in, what time of year it is and whether something is happening nearby. Get those four right and the forecast will usually be close. Many of the other things people point to, from the mood of the team to how many likes the last post got, explain almost nothing.

This guide covers what demand forecasting means in a restaurant, the four inputs that explain most of the week, the inputs that explain little, how to put them together in a few minutes and how to turn the forecast into prep, staffing and a sales figure. Every venue and figure in it is an example.

Demand forecasting is estimating how much customers will want of something in a coming period. In a restaurant, it means estimating how many guests will come for each service in the days ahead: lunch and dinner, each day of the coming week.

That estimate drives almost every practical decision: how much to prep, how many staff to roster, how much to order from suppliers and which tables to hold for walk-ins. A good forecast makes all of them easier. A poor one, or none, leads to waste on quiet nights and shortages on busy ones.

Large businesses use statistical models for forecasting, but most of what those models find in a restaurant's numbers can be seen with a simple table and a few minutes each week. The patterns are strong: Saturdays are busier than Tuesdays, December is busier than January and a booked-out night is busier than an empty diary.

A venue that looks at the same few inputs each week, writes down its forecast and checks it afterwards will usually forecast well enough to cut waste and staff sensibly. It does not need software beyond its own records and a sheet of paper or a spreadsheet.

Demand forecasting estimates guests; sales forecasting estimates the money they will spend. The two are linked: multiply forecast covers by the expected spend per cover and you have a sales forecast. Most restaurants find it easier to forecast covers first, because covers follow clearer patterns than spend.

Forecast covers for each service, then turn them into sales. Our guide to sales forecasting for next week, in twenty minutes covers the routine for doing both each week.

Four inputs explain most of the variation in a restaurant's week:

Together they usually account for most of the difference between one service and another. The sections below take each in turn.

The best single guide to how a service will go is how the same service went in recent weeks. Tuesday dinner tends to look like other Tuesday dinners; Saturday lunch like other Saturday lunches. The day of the week and the service explain more of a restaurant's variation than anything else.

Look at the same service over the last four weeks. If Tuesday dinners did 42, 38, 45 and 41 covers, a forecast of around 40 is a sensible starting point, before the other inputs adjust it.

Recent weeks usually matter more than older ones, because they reflect the venue as it is now: its current menu, its current reputation, the current season. A simple average of the last four weeks works well, and some venues give the most recent week a little more weight.

Watch for trends. If the same service has been rising steadily for a month, the next one is likely to be a little higher still. If one week was unusual, such as a storm or a big event, leave it out of the average rather than letting it skew the forecast.

Bookings already made are the second input, and for many venues the most precise. If Friday already has 70 covers booked by Monday, the forecast for Friday should start from 70, not from an average.

The question is how many more will come. Bookings keep arriving between now and the night, and walk-ins add more on the day. A venue that knows its usual pattern can add those to what is already booked.

The extra bookings that arrive between the day you forecast and the night itself are sometimes called pick-up. Most venues find a steady pattern: perhaps a Friday that has 70 covers booked on Monday usually ends with about 90 booked, because guests book later in the week.

Work out your own pick-up by noting, for a few weeks, how many covers were booked on the day you forecast and how many were booked by the night. The difference, service by service, is your pick-up. Add it to this week's bookings, and the forecast becomes much more accurate.

Walk-ins are guests who arrive without booking, and they vary by venue and by service. A city-centre lunch may be half walk-ins; a destination dinner may have almost none. Knowing the usual share for each service is part of the forecast.

Count walk-ins for each service over a few weeks, and you will see the pattern. Add the usual number to the booked covers and the pick-up. On nights when the book is nearly full, there will be fewer walk-ins, because there is less room for them.

Seasonality, the way trade rises and falls through the year, is the third input. Most restaurants are busier in December than in January, busier when the terrace is open, quieter during the school holidays or busier during them, depending on the venue.

Recent weeks already reflect the current season, but they can be slow to notice a turn, such as the start of the festive season or the quiet after it. Looking at how the same weeks went last year shows when the change usually comes. Our guide to seasonality: planning for the quiet eight weeks covers mapping your own year.

Comparing with the same week last year is useful, but needs care. Calendar dates fall on different days each year, holidays move, and the venue itself may have changed: a new menu, a refurbishment, a new competitor nearby. Last year's number is a guide, not a forecast.

Compare the same weekday in the same part of the year, rather than the same date, and adjust for how the venue has been trading recently compared with last year. If this year has been running about ten per cent busier, last year's figure plus ten per cent is a reasonable seasonal guide.

Local events and holidays are the fourth input, and the one most often forgotten. A football match on the big screen, a concert nearby, a conference in town, a school holiday or a public holiday can change a service far more than any normal pattern.

Keep a simple calendar of events that affect your venue, and check it every week when you forecast. Our guide to local events: the match night nobody told the kitchen covers how to build one and which events matter.

Some things are mentioned often and explain very little: a feeling that "it will be busy", how many likes the last social post got, national news about the restaurant industry, what a competitor did last weekend and the mood of the team. They make for conversation, not forecasts.

Leave them out. If a hunch turns out to be right, look for the real reason, such as an event you had not noted, and add that to the inputs. Over time, the forecast should rest on things you can see and check.

Weather sits in between. For venues with a large terrace or a seaside location, a sunny weekend can make a big difference. For most indoor restaurants, the effect is smaller than people think: very bad weather can keep some guests at home and bring others in from the street, and the two often roughly cancel out.

Note the weather when you check your forecast afterwards. If a pattern appears, such as a terrace that doubles lunch covers on sunny days, add a weather adjustment for that service. If no pattern appears, stop blaming the weather.

A simple method: start with the average of the same service in recent weeks. Look at the bookings already made, add the usual pick-up and walk-ins, and use whichever figure is higher as the starting point. Then adjust for the season, using last year as a guide, and for any events in the calendar.

Write the forecast down for each service, with a note of the reasons for any adjustment: "Saturday dinner 110, up from 95, because of the concert nearby". It takes a few minutes for a week, and the notes make the next forecast easier.

Forecast each service separately: lunch and dinner, and any other service such as breakfast or a late menu. They follow different patterns, are driven by different guests and need different prep and staff. A single daily figure hides those differences.

For venues with very different parts of the room, such as a bar and a dining room, it can help to forecast them separately too. The more closely the forecast matches how the venue actually runs, the more useful it is.

Once you have covers, multiply by the expected spend per cover for that service to get a sales forecast. Spend per cover varies by service and by day: a Saturday dinner may average more than a Tuesday lunch. Use the recent average for each service.

Adjust spend for anything that will change it: a set menu at a fixed price, a special event with a different menu, a large party booked on a fixed package. Sales forecasting built from covers and spend in this way is usually more accurate than guessing a sales figure directly.

The forecast earns its keep when it drives decisions. The kitchen uses forecast covers, and the share of each dish they usually order, to decide how much to prep. The manager uses them to decide how many people to roster for each service.

Our guides to prep from forecast, not from yesterday's prep and scheduling from forecast, not from memory cover how to turn a forecast into a prep list and a rota.

A forecast is only useful if it is reasonably accurate, and the only way to know is to check. After each week, compare the forecast for each service with what actually happened, and note the difference. Over a few weeks, you will see whether the forecast tends to be too high, too low or about right.

Measuring forecast accuracy also shows which inputs need more attention: perhaps the pick-up is underestimated on Fridays, or walk-ins are overestimated at lunch. Our guide to forecast accuracy: measure it or stop forecasting covers how to score it.

A forecast that takes hours will not be done every week. Keep it to a short routine at the same time each week, such as Monday morning: open last week's numbers, the bookings for the week ahead, last year's same weeks and the events calendar, and write down a figure for each service.

With practice, it takes about twenty minutes. The routine matters more than the method: a simple forecast done every week, and checked, beats a sophisticated one done occasionally.

Every forecast is wrong sometimes. When a service is much busier or quieter than forecast, find out why. Was there an event you missed? Did a large party book late? Was the weather unusual? Did a competitor close or open?

Write the reason down. Often the answer is an input you can add next time, such as an event to put in the calendar or a pick-up pattern you had not noticed. A forecast improves by learning from its misses.

A new venue has no history, so the first input, recent weeks, is missing. Rely more on bookings already made, and on what you know from similar venues nearby, until a few weeks of your own numbers build up. Forecast cautiously at first, and check daily.

After a month or two, the venue's own patterns appear and the forecast can follow the normal method. Keep the early forecasts and their results, because they show how quickly the venue settled into its rhythm.

In a group of venues, forecast each venue separately. Each has its own patterns, guests, events and seasons. A group-wide figure is useful for overall planning, but prep and staffing happen venue by venue and need each venue's own forecast.

Sharing the method across the group helps: the same inputs, the same routine, the same way of checking accuracy. Managers can then compare notes on what affects their venues and learn from each other's misses.

The most valuable part of a forecast is often the notes beside the numbers: why Saturday was raised, why Wednesday was lowered, which event was expected to make a difference. When you check the forecast afterwards, the notes show whether your reasons were right.

Over a year, those notes become a record of what drives your venue's trade, which no general model could give you. They are the reason a simple forecast, done carefully, can beat a complicated one.

An invented bistro has been guessing how busy each week will be and regularly throwing away prepped food on quiet nights. The manager starts a twenty-minute forecast every Monday. For Friday dinner, the last four Fridays averaged 92 covers; 68 are already booked, and Fridays usually pick up about 25 more by the night, plus around 8 walk-ins. That suggests about 101.

The same week last year was a little quieter, but the bistro has been running busier this year, and there is a concert nearby on Friday, so the manager forecasts 108 and notes why. Friday does 104. Over the next month, the kitchen preps from the forecast, waste falls noticeably and the manager adds the concert hall's calendar to her list of local events.

servearound does not make forecasts: there is no forecast in its reports, so the sums in this guide are yours. What it keeps are the inputs. Reservation analytics show covers day by day, by meal period and by day and hour, with each headline figure against the period before, how much notice guests give and bookings by source, including walk-ins, so the same service in recent weeks, last year's season and the share of walk-ins are all there.

Bookings already made for the days ahead show on the reservations screen, date by date, and the dashboard shows covers and bookings today alongside revenue and covers against the previous period. servearound keeps no calendar of local events and no weather, so those inputs come from your own notes. See Analytics.

Estimating how many guests will come for each service in the days ahead.

No. Your own records, a few inputs and a simple routine are enough.

The same service in recent weeks, bookings already made, the time of year and known events.

The extra bookings that arrive between the day you forecast and the night itself.

For terraces and seaside venues, often. For most indoor restaurants, less than people think.

Covers first, then multiply by spend per cover for sales.

About twenty minutes a week, once the routine is set.

Compare it with what happened each week and look at the pattern of misses.

Demand forecasting does not need a data scientist. Most of the variation in a restaurant's week comes from four inputs: the same service in recent weeks, bookings already made plus the usual pick-up and walk-ins, the time of year with last year as a guide, and known local events and holidays. Gut feelings, social media likes and national news explain almost nothing, and weather matters mainly for terraces.

Start from the recent average or the booked covers plus pick-up, adjust for the season and events, and write down a forecast for each service with the reasons. Turn covers into sales, prep and staffing, check forecast accuracy every week and learn from the misses. A twenty-minute routine done every week beats any model used occasionally.