Your forecasting model is probably graded on the wrong test.
Most retail demand reporting rolls everything into one accuracy number. Center store, household goods, canned and boxed and bottled, all averaged together into a figure that looks respectable in a steering committee deck. That number is comfortable. It is also useless.
Here is why. Shelf-stable products forgive your mistakes. Overforecast canned tomatoes and the error quietly becomes inventory. It sits. It sells in two weeks, or three, or next season. The miss never surfaces as a line item anyone owns. It gets amortized into carrying cost and disappears.
Perishables do not forgive. Overforecast berries, rotisserie chicken, bagged salad, or fresh fish and the error does not hide. It rots. Within 48 to 72 hours your model’s miss is sitting in a dumpster with a price tag on it. That is not a knock on fresh. That is the gift of fresh. It is the only category that returns a fast, honest, unforgiving signal about whether your demand model actually works.
Fresh is the only honest scoreboard
The scale of the problem is not subtle. Fast Company recently covered what it called America’s 27 billion dollar grocery waste problem, and the reporting makes clear this is now a core operations question, not a sustainability footnote. Afresh is now in use in more than 12,500 grocery store departments nationally, including Safeway and Albertsons.
The most important sentence in that whole piece was not about waste tonnage. It was about accuracy compounding. By better predicting how much can sell in the store, it helps reduce waste in other parts of the supply chain, and when you clean up store ordering, it makes it easier for distribution centers to buy the right amount.
Read that twice. The fresh forecast is not just grading itself. It is grading everything upstream of it. Fix the shelf and you fix the DC. The signal that starts in produce flows backward through the entire network. That is the opposite of center store, where errors flow nowhere and teach you nothing.
Industry data backs the stakes. In 2024, retailers generated 3.98M tons of surplus food, 25% of which went to landfill or was incinerated as waste, with most of it coming from produce, dairy and eggs, dry goods, and fresh meat and seafood, and nearly half caused by concerns or confusion over freshness date labels.
Produce alone is a third of the loss. If your model cannot nail produce, it cannot nail anything that matters.
Why most models quietly fail the fresh test
Fresh is hard on purpose. Retail establishments encounter unique challenges in managing perishable inventory, where the temporal constraints of product shelf-life intersect with volatile consumer demand patterns, and they operate within narrow margins while attempting to maintain product availability and freshness standards that consumers expect.
It gets worse at the clock level, not just the day level. The academic work on ultra-fresh is blunt about this. Perishables require two types of forecasts: one for the first filling of the shelves, and a second for the multiple replenishment during the day, matching expected demand for the remaining opening hours with the fresh inventory still available, which becomes further complicated as products perish at different rates and customers have varying perceptions of freshness.
A model that forecasts in daily buckets will never see this. The error lives inside the day. By the time your weekly accuracy report runs, the signal is already in the trash.
And the legacy failure mode is predictable. Many retailers still rely on outdated manual processes in which store managers respond to stockouts by overordering, and this cycle of overcorrection is a documented driver of excess inventory and eventual spoilage. Your model does not just need to be smart. It needs to replace a human overcorrection loop that has been rewarded for years.
Make fresh your Cost of Doing Nothing line
Here is the forward-deployed move. Stop measuring blended accuracy. Build a perishables-only scoreboard and put the Cost of Doing Nothing in dollars next to every point of forecast error. Not tons. Not CO2. Dollars of fresh margin walking into the compactor every single day you do not act.
The CODN in fresh is the cleanest number in your P&L because it settles in days, not quarters. That speed is your advantage, not your enemy. A model that is graded weekly on shelf-stable goods learns slowly and lies often. A model graded daily on berries learns fast and cannot lie at all.
Chief Data Officers keep asking whether their forecasting investment is working. The answer is already sitting in the back of every store at closing time. Go look at the fresh shrink report. That is your real accuracy metric. Everything else is a rounding error you chose to believe.
Grade the model where it bleeds. Fresh is the test. The waste bin is the answer key.