Demand forecasting errors put retailers at risk of excess stock

AI can help retailers identify patterns in product demand.

Retailers risk holding excess inventory when forecasts fail to accurately predict demand, particularly for new products with limited sales history.

An unnamed beauty retailer's forecasts for new products were almost twice actual demand, resulting in slow-selling products and obsolete stock, Ian Yong, senior consultant at Wavestone Asia, said during the Retail Asia Summit 2026 in Hong Kong.

He said forecasting new products is challenging for beauty retailers because they regularly introduce new items, making it difficult for merchandisers to estimate demand based on historical sales.

To address this, Yong said the team used artificial intelligence (AI) to identify comparable products and analyse their characteristics, including product type, market relevance, price tiers and seasonality.

“The model generated a demand curve for the new product to help the retailer determine what to buy, when to buy it, where to place products and how much to order,” he added.

He also said AI can help retailers identify patterns in product demand and determine when seasonal demand or interest in a new product begins to decline.

Beyond forecasting, Jeanne Pigassou, senior manager and AI lead at Wavestone, said retailers have traditionally relied on merchandisers' experience and intuition to decide which products should be allocated to individual stores.

“This approach can introduce bias into decisions and leave companies exposed when experienced merchandisers retire or move to another business,” Yong said.

Yong also cited a project involving fashion retailer Zadig & Voltaire, where store clustering using a K-means algorithm improved sell-through performance by 10 points.

“The model used 28 numerical features, including customer profiles, store environments and store sizes,” Yong said.

He added that other variables could include average air-conditioning temperatures and the time needed to walk to the nearest train station.

“Ultimately, there's still a human in the loop who helps us to understand what we should do,” he said.

Yong said retailers face different forecasting challenges depending on whether they use push-based or pull-based supply chain models.

In push-based models, retailers decide how much stock to allocate to stores, often without historical demand data for new products. “In pull-based models, customer purchases generate inventory signals that guide replenishment,” he said.

Despite the differences, Pigassou said both models require retailers to balance product availability and inventory levels.

Yong also noted that some retailers continue to calculate safety stock and inventory signals manually on a monthly or quarterly basis, despite having access to large amounts of data.

Looking ahead, Yong said AI could take on more repetitive merchandising tasks, including processing customer orders and generating purchase orders.

This would allow merchandisers to spend more time creating and orchestrating demand, developing business rules and managing exceptions, he said.