Measure forecast accuracy in profit, not in percentage points

How much an algorithm improves the forecast

The first question: by how many percent can a modern algorithm realistically improve the forecast compared with standard statistical models, and what does that do to tied‑up capital and waste?

I wouldn't give a single number, because accuracy doesn't depend on the model alone. Clean, high‑quality input data decides a lot, and the differences can be large even between categories within one company. The typical improvement is somewhere between 5 and 20%. More useful than accuracy percentages is the translation into profit: for a manufacturer, one extra point of accuracy means roughly 0.2–0.4% more profit; for a retailer, 0.1–0.2%.

A supplier outage doesn't belong in the data as it happened

The second question: apart from sales history and the promo calendar, which external data matter most for accuracy, and how does the system handle extremes such as supplier outages or sudden demand spikes?

There are many relevant inputs: in‑store display, item relationships within a category, meaning substitutes and complements, price changes, and the behaviour of the market and competitors. With extremes, it's mainly about how you work with history. A supplier outage or a one‑off demand spike has to be cleaned out of the data or adjusted to normalised sales, and the method depends on the cause, because an empty shelf, a competitor's promo and a data error each need different treatment. In our Veritico STOCK this is a standard step before the forecast itself, not a manual fix when someone spots an odd number.

When a person should still approve the order

The third question was about the line between full automation and human judgement: when should the system leave the final approval of an order to a store manager or buyer?

Where there's enough data, today's models can manage on their own. What matters more is whether the system handles physical constraints on top of the modelling: pack sizes, warehouse capacity, transport. If it does, manual intervention at store level stops paying off, because decisions are made per store, so the labour cost is high and the benefit is small. Some validation makes sense in central purchasing for distribution centres, where strategic goals come into play. Even there, though, the strategy belongs in the model, and the buyer should be able to simulate several scenarios and pick the best one. That scenario capability is exactly what we have in Veritico.

Where this appeared

Algoritmy střídají dohadování (Algorithms are replacing guesswork), Pavel Gregor, Retail News 10/2026, pp. 36–37 (print edition, in Czech)