AI demand forecasting: more accurate than we promised
- 2 days ago
- 5 min read
By: Roni Kleme, AI Engineer
An accurate demand forecast is the heart of Supermind's automated replenishment. In our January product news we estimated that our new AI-powered forecasting models would reach over 85 % accuracy, and AI forecasts are already in use with several customers as part of their daily replenishment and purchasing decisions. Forecast accuracy has now been measured on a customer's real sales across their entire chain: every figure in this article is based on the sales of tens of thousands of product locations. The promise held, with room to spare.
The results: forecast accuracy at location and chain level
The AI's accuracy is 95–96 % at the chain level and around 88 % at the level of an individual location. In January we promised over 85 %.
Our current statistical models have served customers reliably for years, and they set a high bar for the comparison. Even so, the AI improved accuracy further on every period: by nearly four percentage points on average at the chain level.
The clearest gain was measured on A-class products, where the improvement averaged nearly four and peaked at over six percentage points. A-class products are the top tier of an ABC classification: the best-selling items that together account for about 80 percent of total sales. They are just under a third of the assortment but the bulk of the order lines, so the improvement lands exactly where it matters most. The advantage was not limited to the top tier: on B-class products, too, the AI was on average more accurate than the current forecasts, and together the A and B classes cover about 95 percent of sales.
In percentage points the gap looks modest, but the chain-level forecast error fell from 8.2 percent to 4.6 percent: over 40 percent of the error disappeared. At the scale of a whole chain, that is a large amount of money saved. Nor is this just over- and under forecasts cancelling out in the total: the AI forecast predicts what is actually sold more accurately than the statistical models.
How the figures were measured
We compared the AI's weekly forecasts against actual sales on three separate seven week test periods during spring and early summer 2026. As the point of comparison we used the customer's current statistical forecasts.
The forecasts were calculated at the most granular level possible, per product-location. When we report location or chain-level results, they are always sums of these thousands of individual forecasts.
Accuracy was measured one week at a time by comparing the forecast against that week's actual sales. For example, 95 % accuracy means that over the seven-week period, the forecast differed from actual sales by a total of just 5 %.
Where does the AI's advantage come from?
The AI learns from the chain's entire sales history at once and follows no single rule. Instead it combines several signals, such as:
Sales history over short and long horizons
Sales profile: how often and how much a product typically sells
Trend: whether a product's sales are rising or falling
Future information, such as seasons, campaigns, price changes and holidays, which are known in advance from the calendar
Future information is one of the AI's greatest strengths, because statistical methods cannot exploit it: they can only extend history.
Campaigns: future information is a forecast's most valuable ingredient
We measured the effect of future information with campaign data, and it showed up in three ways:
Starting campaigns: no forecast can anticipate a campaign it does not know about. When the model was told about upcoming campaigns, the forecast accuracy of these products rose by 28 percentage points, reaching 97 percent.
Non-campaign products: forecasts for regular products became more accurate too, because the model learns to separate campaign spikes from normal demand.
Ordering: shortages and excess stock decreased at the same time, each by nearly a tenth.
The biggest benefit comes when the model is also told about upcoming campaigns. That is why we are connecting the campaign calendar to the forecast together with our customers.
How good is 95–96 percent accuracy?
A useful yardstick is the M5, the world's largest retail forecasting competition, where over 5,500 teams forecast Walmart's sales. The competition's toughest statistical benchmark model proved so hard to beat that only 7.5 percent of teams managed it.
We ran the same contest on our own data: we used the competition's official metric and calculated the same statistical benchmark model's forecasts on our sales. Our model cleared that bar at every aggregation level, and at chain level our score reached the same level as the competition's winning entry.
The comparison is stacked against us: Walmart's volumes are enormous and the typical production the competition sold nearly every week of the year, whereas Finnish retail assortments are wide and a large share of products sell only part of the year. In other words, forecasting technology thatwould hold its own at the very top of the field is now generating replenishment proposals ineveryday Finnish retail.
More accurate forecasts, smaller safety stocks
Perfect shelf availability would be easy to achieve by always ordering far too much. The real job of a forecast is balance: good availability with the smallest possible inventory and waste. That balance is achieved with safety stocks, and the need for safety stock follows directly from forecast error: the more accurate the forecast, the smaller the buffer needed to secure the same shelf availability. The AI's accuracy advantage does not stay on a report: over time it shows up as lighter inventory across the whole assortment.
Today, safety stocks are typically sized with rules based on the ABC classification, which is a workable way to manage settings for thousands of products at once. The upcoming improvement takes sizing a step further: it is based on each product-location's measured forecast accuracy, complemented by rules that handle exceptions. Accuracy information will also guide the choice of forecasting method going forward. There is no need to wait for the AI forecasts: they are in use today, and the updates will complement them. We will share more closer to release.
Questions about the AI behind the forecasts?
Reach out — happy to answer anything about the models or the forecasting approach.
Roni Kleme
AI Engineer, Supermind
If you want to know what AI forecasts would mean for your assortment and your ordering, get in touch through the contact form or email petri.lindholm@supermind.com directly.
