How Accurate Does a Prep Forecast Need to Be?

By Matt Wampler, CEO of ClearCOGS

Quick Answer: There is no single accuracy number that makes a prep forecast good. Accuracy measures how close a prediction lands to actual demand, but it treats running out and throwing away as the same size mistake, and in a kitchen they never are. The more useful target is a deliberate decision, made item by item, about how often you are willing to run out. For a signature item that brings people in, the answer might be almost never. For a cheap side with a two-day shelf life, running out occasionally is the cheaper outcome.

Why the accuracy question comes up first

Every evaluation of a forecasting system reaches the same moment. Someone asks how accurate it is, a percentage gets quoted, and the room nods. Nobody in the room is quite sure what would have counted as a bad answer.

This is understandable. Accuracy sounds like the thing you should care about. It is a single number, it is comparable across vendors, and it feels objective.

It is also close to meaningless on its own, and building your expectations around it leads to worse operating decisions than asking a harder question.

What does forecast accuracy actually measure?

Forecast accuracy measures the distance between what a model predicted and what actually sold. Accuracy figures usually come from an error measure such as mean absolute percentage error, reported as one hundred percent minus the error.

Two things about that are worth knowing before you use the number.

It depends entirely on what is being measured. Accuracy on total daily sales for a location is a much easier problem than accuracy on a single ingredient for a single daypart. A system reporting a high figure at the store level and a system reporting a lower figure at the ingredient level may be doing exactly the same quality of work on very different questions. Always ask what level the number describes.

It treats both directions of error as equivalent. Being twenty portions over and twenty portions under produce the same accuracy score. In your kitchen they produce completely different outcomes, and that is the whole problem.

What is the actual tradeoff?

Every prep decision is a bet placed before you know demand. You will be wrong in one of two directions, and the two are not symmetrical.

Prep too much and you pay the food cost of what gets thrown away, plus the labor to make it. The loss is real, bounded, and visible in the walk-in at close.

Prep too little and you lose the margin on sales you could have made, and something harder to measure: a guest who came for a specific item and did not get it. Some of those guests shrug. Some of them do not come back, and some tell other people. That cost is invisible in the same way, which is precisely why it gets underweighted.

The ratio between those two costs is what should determine your target, not a generic accuracy percentage. This is an old idea in operations research, formalized in the 1950s as the single-period inventory problem, and the conclusion is direct: the right quantity depends on the relative cost of having too much versus too little, which sets how often you should expect to run short.

How do you set the target item by item?

The practical version does not require math. It requires sorting your menu.

Most operators already know this intuitively. What they usually have not done is write it down and make it a standard, which means every location and every manager is making the call differently.

Two questions get you most of the way there. If we run out of this at seven o’clock, what actually happens? And if we have twelve portions left at close, what actually happens to them? Items where the first answer is worse than the second should be prepped long. Items where the second answer is worse should be prepped tight.

Why not just leave this to the manager?

Because the evidence says human judgment on exactly this decision is biased in a predictable direction, and experience does not correct it.

Researchers ran experiments on how people set order quantities under uncertainty with known costs. When the cost of running out was larger than the cost of excess, participants ordered more than the optimal amount, behaving in a risk-seeking way. When excess cost was the larger of the two, they ordered less than optimal. In both cases the deviation was systematic rather than random (Surti, Celani, and Gajpal, European Journal of Operational Research, 2020).

The finding that matters most for operators is what happened next. The researchers supplied performance feedback, including explicit market information, to see whether people would correct. They did not. Feedback produced no meaningful reduction in the bias, and over repeated rounds participants drifted further from the neutral optimum rather than closer.

That should change how you read a familiar situation. When a location consistently over-prepares one category and under-prepares another, the instinct is to treat it as a training gap or an attention problem. It may just be what happens when a person makes a repeated decision under uncertainty without a reference point. More reporting alone is unlikely to fix it.

So what number should you hold a system to?

Ask for four things instead of one.

1. Accuracy at the level you care about. Ingredient or item level for a specific daypart, not store-level daily sales.
2. The direction of the misses. A system that is wrong evenly in both directions is behaving differently from one that is consistently short, even at identical accuracy.
3. Performance on your top items specifically. Aggregate accuracy can be carried by a long tail of predictable low-volume items while the twenty items that drive your food cost stay noisy.
4. Whether the target can be set per item. If every item is forecast to the same expected value, you have no way to express that running out of one thing is much worse than running out of another.

That fourth point is the one most often missing, and it is the difference between a forecast that produces a number and a system that reflects how your business actually loses money.

Frequently Asked Questions

Is 80 percent forecast accuracy good?

It depends entirely on what is being measured and on what you were doing before. Eighty percent at the ingredient level for a specific daypart is a meaningfully harder result than eighty percent on total daily sales. The relevant comparison is not a benchmark, it is your current process.

Should a restaurant ever plan to run out of something?

For some items, yes. If an item is inexpensive, has a short shelf life, and has a close substitute on the menu, occasionally running out late in the day is usually cheaper than routinely discarding it. That is a deliberate decision, not a failure.

How is this different from setting par levels?

A par is a fixed quantity you hold regardless of what tomorrow looks like. This is about choosing how much risk of running short you accept, then letting the target quantity move with predicted demand.

Does more history make forecasts more accurate?

Up to a point. More history helps a model learn weekly and seasonal patterns. It does not help with genuine novelty such as a new item or an unusual event, and it does not resolve the over versus under question, which is a business decision rather than a modeling one.

Who should own this decision?

Operations, not the vendor and not the model. The relative cost of running out versus throwing away is specific to your brand, your margins, and your guests. A good system lets you express that. It should not decide it for you.

Accuracy is a useful diagnostic and a poor target. It compresses two very different mistakes into one number and hides the direction of the error, which is the part that costs you money.

The better question is one only your operations team can answer: for each item that matters, how often are we willing to run out? Write those answers down, make them a standard rather than a manager’s habit, and then hold your forecasting to whether it hits them.

Let’s Talk

Sources

• Surti, Chirag, Celani, Anthony, and Gajpal, Yuvraj. The Newsvendor Problem: The Role of Prospect Theory and Feedback. European Journal of Operational Research, 287(1), 251–261. November 2020. sciencedirect.com