Forecasting methodology and process for restaurants

As the second in our series on forecasting, we wanted to talk about a potential approach to forecasting restaurant sales, and then look at implementation.


A) How do you come up with a forecast?

At Tenzo, we follow a broad four step process to forecasting restaurant sales:

  1. We use traditional forecasting methodologies to understand the core components of a forecast i.e.
  • Overall growth trend
  • Weekly fluctuations
  • Annual seasonality

As an example — that might give you something like the following:


2. Second we then use machine learning techniques to include other variables. The kind of variables you might think about including here are things like:

  • Weather — temperature, precipitation, sunshine
  • Events — e.g., holidays, local sports events etc.
  • Traffic

3. We then have to solve the problem of having a top-down forecast vs. a bottoms up forecast.

Top-down: Typically a top-down forecast will be used to set corporate plans, and schedule labor against.

Bottoms-up: Typically a bottoms-up forecast will be used to determine item level procurement, or items to prepare at the start of a shift.

4. Lastly — we also think about intraday. Intraday forecasting allows you to adjust throughout the day. Typically there will be limited flexibility in terms of labor (although sometimes you can make small adjustments) but this does allow you adjust prep schedules and try to match the food you are making to

Of course — one of the most important thing is to build a machine that will learn over time. And we’ll talk about that in the next post!

B) How do you then implement this forecast?

To some degree, this varies by business but we recommend 4 core phases:

1.Forecast setting period: During this phase the manager takes all of the information available to him to consider whether to adjust the provided forecast (depending on your technology, this may require more or less adjustment)

2. Forecast approval: Now you need some process for forecasts to be submitted “upwards” i.e. a store manager, submits to HQ, or an area manager. The supervisor then had an opportunity to adjust, or lock-in the provided forecast

3. Forecast action: This is the fun bit — now you’ve got a forecast you have to do something with it, we recommend:

a) Build out a labour schedule for the week that will mean you can serve that level of demand (e.g., have the right number of servers vs. chefs). Critically, if you can do this by hour — you’ll save even more!

b) Adjust procurement volumes based on inventory levels and forecast demand — place orders, or transfer inventory, to make sure you are going to be able to serve expected demand

c) Build any necessary prep sheets to ensure you are starting the day off right

4. Adjustment: A critical fourth phase is the dynamic adjustment phase. Things will happens throughout the week (e.g. dramatic change in weather forecast, the cancellation of a sports event), that will mean you need to adjust. This is worth looking at every day for extreme variances. If you notice you’re off, we’d recommend:

a) Adjusting daily prep sheets immediately

b) Filling a labor gap with short term staff is possible

c) Run flash-sales, or tweak menus to deal with short-term under/over supply of inventory

We hope it’s helpful — please let us know your thoughts. We’d love to help you get your numbers as accurate as possible.

Next time we’ll go in to how your forecast can learn over time.

If you’re interested in a demo — sign up here.

How does weather affect restaurant sales?

This is the first of what is intended to be a 3 part series on how to get quality forecasts. We’d love your feedback!

At Tenzo we’ve spent a lot of time thinking about how to forecast accurately and wanted to share a couple of the key insights. Typically we have found:

  1. A smart forecasting algorithm can out-perform a store manager in the long run by 25–50%, but it might miss short term events (e.g., roadworks, staff shortage) that only a manager may know about
  2. Weather and events are a critical element in improving a typical forecast
  3. Rain affects sales, but after a certain amount of rain the impact diminishes. Extreme temperatures (vs. normal for the season) are what cause significant deviations in sales.

How do we think about the impact of machine learning?

The below chart compares a 4-week rolling average (e.g., the last 4 Mondays) to a machine learning generated forecast. We typically find a 4-week average this is the best a good manager can do given their memory for events.

MAPE, for a 4-week average forecast vs. a machine learning generated forecast


As you can see — the computer wins overall. There may be days when a manager knows something the computer does not, but overall results tend to this.

Now, let’s look at a chart showing how rain impacts sales for a given location.:

Daily sales variance per mm of rain


The result is clear — when it start to rain, sales drop-off, but beyond a certain point — people are no longer driven by the rain.

Now, let’s look at temperature:

Daily sales variance per degree of temperature (celsius)


Interestingly — this chart is looking at a day in July. A big portion of temperature will be captured in the normal seasonality (which any computer based forecasting system should capture). However, what becomes apparent is how it’s all about extremes — if the day is unseasonably warm or cold it will dramatically affect sales up to a point.

Note: importantly all of these results will be dependent on the specific location, brand and type of business. That’s why you need a machine!

Thanks for reading, we hope we’ve shown:

  1. How weather can be an important factor for your business
  2. The kind of dramatic improvement you can get by looking at it
  3. How a computer can outperform a typical location manager

Let us know your thoughts!

To arrange a demo of Tenzo, please visit