By Matt Wampler, CEO of ClearCOGS
A technology leader at a national entertainment and dining operator recently described what happened when his team switched on the sales projections module in their back-office system. The projections came out more than 40 percent above actual sales. Not on a bad week. Consistently. His summary was blunt: it was a straight line through a business that does not move in straight lines.
The result was predictable. Nobody trusted the number. His venue leaders assemble the real forecast by hand, pulling history from one system and booked events from another into a spreadsheet, and both his back-office and labor platforms get manually overridden every week. He is paying for forecasting twice: once for the software, and again in the salaried hours spent correcting it.
Here is the part most operators miss. The system was not bad at math. It was doing math on the wrong question.
A Forecast That Predicts Everything Predicts Nothing
This operator’s revenue is a blend. There is walk-in traffic, which is genuinely uncertain. There are corporate events, which are booked weeks in advance and sit in a system already. There are hybrid bookings, partially reserved and partially day-of. And there is retail, which drives no prep and no kitchen labor at all.
His back-office system forecast the whole blend as one number. That is statistically doomed. Every large booked event lands in the history as a spike, and a model that cannot tell a booked event from organic demand dutifully extrapolates those spikes forward into weeks that have no events on the calendar. The output runs structurally high, which is exactly what he observed. The same visit slot at his venues can be worth thirty dollars on a quiet weekday morning and four hundred dollars on a Friday night; averaging across that without knowing why produces a number that describes neither.
The failure is not the model. It is the scope. The system was re-predicting revenue that was already sitting, fully known, in another database.
The Override Trap
The rational response to a bad forecast is to override it, and that is what his teams do, weekly. But override culture has a well-documented cost. A landmark study in the International Journal of Forecasting examined more than 60,000 real forecasts across four companies and found that small manual adjustments often damaged accuracy, and that upward adjustments in particular were much less likely to improve the forecast, reflecting what the authors called a general bias toward optimism.
This operator has lived that finding. Earlier in his career at a national chain, his team ran the experiment properly: a control group of stores that left the model’s numbers alone, against stores that adjusted them. The stores that touched the numbers did worse. Every single time.
So the trap closes from both sides. A forecast that predicts everything is wrong, which forces overrides. And routine overrides make the numbers worse while training everyone to distrust the system that produces them. Once a forecast is overridden every week, you do not have a forecasting system. You have a suggestion box, plus an unwritten process living in one leader’s spreadsheet. His question about that spreadsheet is the one every operator should ask: what happens when that person leaves?
Forecast Only the Unknown
The fix is decomposition. Before asking how to forecast better, sort every revenue stream into one of three buckets.
| Revenue stream | Is it known in advance? | What to do with it |
|---|---|---|
| Booked events and catering | Yes, it is in the events system | Subtract it from the forecast. Feed it to prep and labor directly. |
| Hybrid or partial bookings | Partially | Take the booked portion as known. Forecast only the day-of remainder. |
| Retail and non-kitchen sales | Irrelevant to prep and labor | Exclude it. A number that drives no decision is noise. |
| Walk-in food and beverage demand | No | This is the forecast. Spend all the statistical effort here. |
Assumptions: your booked events live in a queryable system and your point of sale separates revenue centers. Most modern stacks do both.
Two things happen when the forecast is scoped this way. First, the statistical problem gets dramatically easier, because the model is no longer trying to predict spikes that were never random. Second, trust becomes possible, because the number now means something specific that a kitchen manager can act on: this is the walk-in demand we expect you to prep for, on top of the events you can already see.
One refinement from this operator’s world is worth stealing: even “known” events are not fully known on the food side. His kitchens routinely send out a third to half of the ordered food for an event, gauge consumption, and hold the rest. Sold revenue and consumed product are different numbers, and ordering to the sold number quietly stocks the shelves with product that will never be served. Known revenue still deserves a consumption model, just not a demand forecast.
The Bar Is Exception-Based
Asked what the future of his operation should look like, this operator did not say better reports. He said everything from projections to recommended ordering should be automated to the point where humans only handle exceptions: the system produces the numbers, a manager glances at them, flags the one thing it could not have known, and moves on. That is the correct bar, and the research supports it. Reserve human judgment for the large, information-driven exceptions where the study found adjustments actually help, and stop spending it on weekly corrections of a number that was scoped wrong to begin with.
This is the layer we work in at ClearCOGS: separating what is already known from what genuinely needs predicting, then forecasting the unknown part per item, per location, per hour, so the number that reaches the kitchen is one the team can stop second-guessing.
The diagnostic for your own operation takes one meeting: list every revenue stream your forecast currently blends, and mark which ones are already sitting in a booking system, a catering calendar, or a retail report. If your system is re-predicting revenue you already know, its accuracy problem is not a modeling problem. It is a scoping problem, and scoping problems are fixable this quarter.
Sources
- Fildes, R., Goodwin, P., Lawrence, M., and Nikolopoulos, K. Effective Forecasting and Judgmental Adjustments: An Empirical Evaluation and Strategies for Improvement in Supply-Chain Planning. International Journal of Forecasting, 25(1), 3-23. 2009. research.lancaster-university.uk
