By Matt Wampler, CEO of ClearCOGS
Here is an uncomfortable property of the most common forecasting method in restaurant back-of-house software: the better your business performs, the worse the forecast gets.
Not noisier. Worse in one direction, every single week.
Most legacy back-of-house tools forecast labor and product demand the same way: take the last three weeks of sales, average them, and project that forward. It is simple, it is explainable, and for a flat business it is roughly fine. But growing restaurant groups are not flat businesses, and the math of a trailing average punishes them for it.
The Two-Week Lag Hiding Inside a Three-Week Average
A trailing three-week average is not an estimate of next week. It is an estimate of the recent past. Averaging weeks one, two, and three ago centers your forecast on demand as it existed roughly two weeks ago.
For a flat business, demand two weeks ago and demand next week are the same number, so nobody notices. For a growing business, they are not the same number, and the gap is not random error. It is bias: a forecast that is low every week, by a predictable amount, in the same direction.
The size of that built-in miss is a function of one variable: your growth rate.
Built-in weekly bias ≈ 2 × weekly growth rate
| Annual sales growth | Weekly growth rate | Built-in under-forecast, every week |
|---|---|---|
| 10% | 0.18% | ~0.4% |
| 20% | 0.35% | ~0.7% |
| 40% | 0.65% | ~1.3% |
These are illustrative scenarios, not benchmarks. A group growing 40% a year is scheduling labor and ordering product against a forecast that starts roughly 1.3% short before anyone makes a single decision. Every week. Compounding as long as the growth continues.
That may sound small. It is not, for two reasons.
Reason One: The Bias Lands on Your Largest Cost Line
Labor is the single biggest line on most restaurant P&Ls. According to the National Restaurant Association’s 2025 Restaurant Operations Data Abstract, salaries, wages, and benefits ran a median of 31.7% of sales for limited-service operators and 36.5% for full-service operators in 2024, both well above their historical averages.
When the demand forecast feeding your labor schedule is systematically low, one of two things happens, and both cost money:
- The schedule follows the forecast. The store is structurally understaffed on its busiest weeks, throughput suffers exactly when the most revenue is on the line, and service times stretch during the rushes that new customers experience first.
- Managers stop trusting the forecast and override it. Now every location is staffing on gut feel, the variance between stores widens, and the forecasting tool has become an expensive report nobody reads.
The second failure mode is the more expensive one, because it is invisible. The tool still technically works. It just stopped being used.
Reason Two: The Bias Peaks on Exactly the Wrong Days
A trailing average has no input for the future. It cannot see the marketing calendar, the stadium event, the festival weekend, or the promotion your marketing team is about to run. Those are precisely the days when demand departs furthest from the trailing three weeks.
| Day type | Forecast miss (40% growth scenario) |
|---|---|
| Ordinary week | ~1.3% low |
| Week with a 10% promotion lift | ~11.3% low |
| Event weekend with a 30% lift | ~31.3% low |
On a promotion day, the model misses the entire lift, plus the baseline growth bias on top. The forecast is at its least accurate on the days when accuracy is worth the most: the days you spent marketing dollars to create.
Operators know this pattern even if they have never written down the math. It is why growing groups develop informal workarounds: nearby stores texting each other to borrow product mid-shift, drivers dispatched from the commissary for emergency replenishment, managers padding prep numbers by feel before a big weekend. Each workaround is a small labor tax paid to compensate for a forecast that was structurally guaranteed to be short.
The Diagnostic: Check Your Error for Direction, Not Just Size
Most operators who review forecast accuracy look at the size of the error. The more revealing question is the direction.
Pull the last 12 weeks of forecast versus actual, at the sales level, per location. Then ask:
- Is the error centered on zero, or is it consistently on one side? Random error is a model quality problem. One-sided error is a model design problem, and no amount of tuning a trailing average fixes it.
- Does the error grow on promotion and event weeks? If yes, your forecast has no forward-looking inputs, and your marketing team is creating demand your operations team is not being told about.
- Are managers overriding the numbers? Track the override rate. A rising override rate is the leading indicator that a forecasting tool is about to become shelfware.
If the answer to the first question is “consistently low,” the finding is worth restating plainly: your forecast error is not noise. It is your own growth, arriving two weeks before your tooling admits it exists.
What a Forward-Looking Forecast Changes
The fix is not a longer trailing average. An eight-week average makes the lag worse, not better. The fix is a forecast built on inputs that lead demand instead of trailing it: the marketing calendar, local events, weather, seasonality, day-of-week patterns, and the growth trend itself, modeled at the item level rather than the sales level.
ClearCOGS works at this layer: turning POS history plus forward-looking signals into daily prep, ordering, and labor recommendations, and then measuring the result against forecast error, waste, and stockouts week over week. The point is not that machine learning is clever. The point is that a model with forward-looking inputs can carry a growth trend and a promotion calendar, and a trailing average mathematically cannot.
Bottom Line
A trailing-average forecast under-predicts a growing restaurant by roughly twice its weekly growth rate, every week, with the miss peaking on promotion and event days. The faster you grow, the larger the built-in error, and the more your teams pay for it in emergency transfers, gut-feel scheduling, and eroded trust in the numbers. Before evaluating any forecasting tool, run one diagnostic: chart the direction of your forecast error for the last 12 weeks. If it lives on one side of zero, the model is not inaccurate. It is pointed backwards.
Want to see what a forward-looking forecast would say about your locations? Let’s Talk
Sources
- National Restaurant Association. Restaurant labor costs are well above historical averages. 2025 Restaurant Operations Data Abstract (2024 data). restaurant.org
