Use casesDemand forecasting
A forecast you can hold to a number, measured against the rule you use today.
Stockouts on one shelf, write-offs on the next. Every planner already has a forecast: last year plus a feeling. We build forecasting that runs in your warehouse, has to beat that rule on months it never saw, and tells you how far it missed by, because a forecast without an error band is a guess with a decimal point.
What gets in the way
A promotion lands and the shelf is empty by Thursday. A season turns and a warehouse is full of what did not sell. Both cost money, and both were decided weeks earlier by a forecast nobody could check.
Most forecasts are last year plus a feeling, and most forecasting tools report an accuracy that was never tested against that simple rule. If a model cannot beat repeating last week, it is not worth the pipeline.
The planners know things the data does not: a customer going on allocation, a plant on maintenance, a storm on Tuesday. A forecast that cannot take that in gets ignored, and should.
What we build
It runs where your sales and inventory data already live, and it has to beat the simplest rule to stay.
- A forecast per product, location and week, rebuilt nightly from your own history, with the calendar, price and promotion signals your planners already track.
- Two baselines run beside it on every refresh: repeat last week, and same week last year. If the model stops beating them for a product, the dashboard says so and falls back.
- An error band on every number. A forecast of 1,200 cases with a band of 1,050 to 1,350 is a decision. A bare 1,200 is not.
- Planner overrides are first-class. A planner adjusts a number, the reason is kept, and the next evaluation shows whether the override helped.
- Native where it can be. Snowflake ML FORECAST and ANOMALY_DETECTION ran inside the account on Grid Intelligence, with the same honest comparison.
How we know it works
Every forecast model we ship is tested out of sample against the rules you already use.
- On Grid Intelligence, our utility platform, the load forecast missed by 8.8% over seven held-out months, against 11.8% for repeating yesterday and 16.3% for last week. It stayed because it beat both. The same test applies to your products.
- The anomaly check that caught a collapsing total in Verdict watches your feed too, so a broken data load is flagged before it becomes next week’s forecast.
- The strongest signal is often not what anyone expected. At Winland Foods, the shipping plant predicted claim outcomes better than anything else. We report what the data says, not what the slide said.
- Accuracy is reported per product and per week, with the baseline beside it, so a planner can see where to trust the model and where to trust themselves.
Built from this work
We have not shipped a demand forecast under a retailer’s or a manufacturer’s name yet. The forecasting discipline and the baselines are live in the work below.
58% of failures caught by a 50-unit watch list on the test fleet
8.8% load forecast error on the test fleet, against 11.8% for yesterday’s load
4 of 4 summer peaks caught in an out-of-sample replay on the test fleet
Read the story28% of claims clear automatically at the recommended bar
92% of cleared claims were real shortages, on months the model never saw
$174K in deductions cleared through the model at the first month-end close
Read the story
Where it fits
Manufacturing
AI for manufacturers, where the count has to be right.
Questions
How accurate will it be?
We will not say until it has run on your history. What we promise is the test: held-out months, two simple baselines, error reported per product and per week, and a fallback when the model loses.
Can it use promotions and price?
Yes, where you have them in data. Each signal is tested for whether it helps out of sample. Signals that only help in hindsight are removed.
What does the planner see?
The number, the band, the baseline and the reason the model moved, in the tool they already plan in. Overrides go back into the evaluation.
How fast is the first forecast?
Winland’s claim model went from kickoff to a production run in 25 days. A forecast with clean history runs on a similar clock. Messy history takes longer, and we tell you where it is messy in week one.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.