6 November 2026International edition
Vol. I · No.
6 November 2026
AI IN FASHION
DAILY
The daily briefing on AI in the fashion business
Where fashion meets artificial intelligence.
Merchandising & Buying · Stories

How to forecast demand at style-colour-size level with AI in 2026

Size curves shift by category and market, requiring SKU-level models that ingest attribute, location and historical sell-through data to avoid markdown waste.

a large number of numbers are arranged in rows
Photo: Enric Moreu / Unsplash

KEY TAKEAWAYS Summary by the editors

  1. Size distribution for the same style varies significantly across product categories, geographies and sales channels.
  2. Modern AI demand forecasting tools model at the individual SKU level rather than aggregating styles.
  3. Effective implementation requires clean historical sell-through data, product attributes and regional size standards.
  4. Organisations typically need six to twelve months of data to train accurate size-level forecasting models.

Forecasting demand at the style-colour-size level allows merchandisers to allocate inventory more precisely, reducing markdowns and stockouts. The challenge lies in modelling the size curve, the distribution of demand across sizes for a given style and colour, which shifts by product category, geography, sales channel and consumer segment. AI tools now model at the SKU level, ingesting historical sell-through data, product attributes and location information to predict which sizes will sell in each market.

Why does the size curve vary by category and geography?

Size curves reflect the body shape and sizing preferences of the target customer base, which differ across product types and regions. A bra size curve, for example, is far more complex than that of a unisex T-shirt, because bra sizing involves two dimensions and a wider range of size options. Within apparel categories, tailored trousers exhibit different size distributions than knitwear, as fit tolerances and customer expectations vary.

Geography introduces further variation. European and North American sizing standards differ, and consumer preferences within those markets are not uniform. A retailer selling in Germany and France may observe that demand for smaller sizes is proportionally higher in France, while demand for larger sizes is stronger in Germany. These patterns are further complicated by shifts in consumer demographics, including ageing populations and changing body measurements over time.

Sales channel also influences size distribution. Online customers may favour different sizes than in-store shoppers, partly because of varying return behaviours and the tendency to order multiple sizes. As a result, merchandisers cannot apply a single size curve across all channels without risking excess inventory in some sizes and lost sales in others.

What data do SKU-level forecasting models require?

AI models that forecast at the SKU level require three main categories of input data. The first is historical sell-through data, which records actual sales by style, colour, size, location and time period. This data should reflect true demand rather than simply sales, meaning that stockout periods must be identified and accounted for to avoid training the model on constrained supply.

The second category is product attributes, including silhouette, fabric, fit type, price point and any relevant design details. These attributes help the model identify similarities between products and infer size curves for new styles that resemble past bestsellers. For example, if a new slim-fit trouser shares fabric and silhouette attributes with previous styles, the model can apply learned patterns about size distribution for that category.

The third category is contextual information such as geography, store format, weather data and calendar events. Location data enables the model to apply region-specific size curves, while weather and calendar inputs help adjust forecasts for seasonal variation and promotional periods. Some advanced systems also incorporate customer demographic data where available, though privacy regulations limit the use of individual-level information.

Which tools model at the SKU level?

Several software platforms now offer SKU-level demand forecasting for fashion, though many merchandisers also build proprietary systems using machine learning frameworks. Commercial tools typically use ensemble models that combine multiple algorithms, including gradient boosting and neural networks, to handle the high dimensionality of SKU-level data. These platforms integrate with existing enterprise resource planning and product lifecycle management systems to pull in historical sales and product attribute data.

Some retailers have developed in-house forecasting systems tailored to their specific assortment and customer base. These custom models allow greater control over feature engineering and the incorporation of proprietary data sources, such as loyalty programme information or customer fit feedback. However, building in-house requires significant data science expertise and ongoing maintenance.

Regardless of whether a tool is commercial or proprietary, effective SKU-level forecasting depends on the quality and granularity of input data. Models perform poorly if historical data is incomplete, if product attributes are inconsistently tagged, or if size standards vary across suppliers without normalisation.

What are the setup requirements for implementation?

Implementing SKU-level demand forecasting begins with data preparation. Organisations must audit their historical sales data to ensure completeness and accuracy, filling gaps where possible and flagging periods of stockout or promotional activity. Product attribute data should be standardised, with consistent taxonomies for category, fit, fabric and other relevant features. This often requires collaboration between merchandising, design and IT teams to establish a shared vocabulary.

Size standardisation is a critical step, particularly for businesses operating across multiple geographies. A medium size in the United States does not correspond to a medium in the United Kingdom or European Union, so the system must map sizes to a common framework or model each market separately. Some organisations create a size index that normalises all sizes to a comparable scale, while others maintain separate models for each regional sizing standard.

Training the model typically requires at least six to twelve months of historical data, though more data improves accuracy, particularly for seasonal categories or slow-moving styles. During the initial phase, merchandisers should run the model in parallel with existing forecasting processes, comparing outputs and refining parameters before relying on AI-generated forecasts for allocation decisions. Ongoing model monitoring is essential, as shifts in consumer behaviour, supplier changes or category expansion can degrade performance over time.

How do merchandisers integrate forecasts into allocation?

Once the model generates SKU-level forecasts, merchandisers use those predictions to inform initial buys and allocate inventory across stores or distribution centres. The forecast provides a recommended size curve for each style-colour combination in each location, which the merchandiser can adjust based on additional business knowledge, such as upcoming marketing campaigns or regional events.

Many organisations adopt a phased allocation approach, placing an initial order based on the forecast and reserving capacity for in-season replenishment. This strategy reduces risk for new styles with uncertain demand, allowing the business to respond to early sell-through data before committing to full inventory depth. For replenishment, the model continuously updates forecasts as new sales data arrives, enabling dynamic reallocation between locations.

Integration with allocation and replenishment systems is essential for operational efficiency. Manual transfer of forecast data between systems introduces errors and delays, so most implementations involve API connections or shared databases that allow real-time data flow. Merchandisers also benefit from dashboards that visualise forecast accuracy by category, location and time period, highlighting areas where the model performs well and where human intervention remains necessary.

What challenges remain in 2026?

Despite advances in AI forecasting, several challenges persist. Data quality remains the most common obstacle, as many fashion businesses still lack comprehensive sell-through data or consistent product attribute tagging. Organisations with fragmented systems, legacy software or frequent supplier changes face particular difficulty in assembling the clean datasets that SKU-level models require.

Another challenge is handling new products with no sales history. While models can infer size curves from similar past styles, truly novel designs or new categories require different approaches, such as test launches in select markets or reliance on broader category averages. Merchandisers must balance the efficiency of algorithmic forecasting with the need for human judgement on innovation and trend-driven products.

Finally, organisational change management is often underestimated. Shifting from experience-based forecasting to data-driven models requires buy-in from merchandisers who may be sceptical of algorithmic recommendations. Successful implementations involve training, transparent communication about model logic and gradual rollout that allows teams to build confidence in the system over time.

Frequently asked questions

How much historical data is needed to forecast at SKU level?

Most organisations require at least six to twelve months of historical sell-through data to train accurate SKU-level forecasting models, though more data improves performance, particularly for seasonal categories or slow-moving styles.

Why do size curves differ between online and store channels?

Online customers often order multiple sizes with the intention of returning some, and the absence of in-person fitting leads to different purchasing behaviour, resulting in size distributions that differ from in-store demand.

Can AI forecast sizes for entirely new product categories?

AI models struggle with truly novel categories that lack sales history. In these cases, merchandisers typically rely on test launches in select markets, broader category averages or similar product analogies until sufficient data accumulates.

Stories