A forecast you can stand behind.

A model that fits your history perfectly has proved nothing; it has only memorised. So Stratos cuts each product’s history once, twelve months back. Every candidate family is trained on everything before the cut and then asked to forecast the twelve months it has never seen — the same twelve months, for all of them, so the comparison is fair.

The families are not variations on one idea. They run from a plain repeat of last year, through moving averages and exponential smoothing, up to methods that can take account of outside factors like Ramadan. Each is fitted to the product on its own, because the rule that suits one product rarely suits the next.

Each month’s gap between forecast and reality is measured, and the families are scored on the total weighted error and on bias, which says whether a model leans high or low rather than merely missing. The winner is the one that actually forecast best, with a preference for models a planner can explain when the margin between them is thin. One competitor is always in the race: the seasonal naïve, which simply repeats last year, and anything that cannot beat that has not earned the product.

Where nothing clears the bar, the item is not quietly passed off as fine — it is flagged on the page, with the gate it failed. And the whole competition is re-run every cycle, so a model that stops working loses its product to another one without anybody having to notice it slipped.

Most tools flag a month because it is far from an average. That catches every Ramadan and misses the quiet disasters. Stratos asks a sharper question: given everything the model knew a month ago, how surprising was what actually arrived? The filter carries a running expectation for each product and a running measure of that product’s own noise, and scores each month’s surprise in units of that noise. The tolerances are deliberately lopsided — a sudden collapse is more often a stockout or a delivery that slipped than a real fall in demand, so it is questioned sooner than an equally large jump. Inside a known event the bar is raised, because a surprise during Ramadan is usually the calendar rather than the market. What breaks tolerance is flagged, with its evidence, and handed to the planner. Nothing is ever corrected on its own.

The Hijri year is about eleven days shorter than the civil one, so Ramadan and both Eids march steadily backwards through the Gregorian months. A model that only knows “March” sees a peak that arrives early, then earlier, then splits across two months — and reads all of it as noise. Stratos places the events themselves on the calendar for every year ahead, along with the weeks that lead into them, the school year and the breaks in it. It then measures, per product, how much each of those actually moves demand: the run-up to Eid matters enormously for gifting lines and hardly at all for a staple. The result is a forecast that expects a February peak to become a January peak, because it knows why.

Stratos offers every product the same list of candidate signals and lets the history decide. Each candidate is measured for how much it moves demand and how confident that measurement is; the ones that cannot earn their place are dropped, and the survivors are shown with their strength, not merely their names. Then comes the part that matters in a review meeting. If a planner believes a signal is missing — or believes one that survived is nonsense — the grid search re-fits the alternatives and ranks them side by side, including the option of using no calendar factors at all. The argument stops being a matter of seniority.

Once the plan is published, Stratos routes it onto the production lines that actually make the products and compares each month’s load against what that line can produce. The result is a single grid: every line, every month, shaded by how hard it is being pushed, with the months that do not fit marked plainly. Where a month overflows, the shortfall is quantified rather than described — this many tonnes, on this line, in this month — and the neighbouring months show their headroom, so the size and the shape of the problem are both visible. What Stratos will not do is decide for you. It does not schedule the factory, it does not move volume between months, and it never quietly trims the forecast to make the arithmetic work. It shows the collision early enough that the people who own the plant can deal with it while there is still time.

Stratos never sizes a buffer from a textbook constant. It collects the shipped model’s real misses over the review window, gathers them into their own distribution, and places each item’s service target on that curve — higher for the A items that carry the business, lower where a miss costs little. The buffer is then capped by the item’s shelf life, because stock that expires is not safety. And the promise is graded in public: the service level the plan committed to, printed next to the service level the shelves actually delivered.