How do size forecasting models aggregate demand across size curves?

Size forecasting models aggregate demand across size curves by pooling historical sales data from similar products, store groups and past seasons to identify stable size distribution patterns, then apply those patterns to forecast demand for new or replenishment inventory. This aggregation overcomes the sparse, noisy data that results from tracking every stock keeping unit at the size level, allowing merchandisers to allocate inventory more precisely and reduce both markdowns and lost sales.

Why is size-level forecasting harder than style-level forecasting?

Forecasting at the size level introduces significant data sparsity. A single style may generate hundreds of transactions across all stores and channels, but each individual size within that style may sell only a handful of units per location, making it difficult to distinguish genuine demand signals from random variation. Stock-outs further distort the picture: if a popular size sells out early in the season, historical sales understate true demand, while slow-moving sizes may appear more popular than they are simply because they remained available longer.

Traditional item-level forecasting models struggle with this sparsity. Retailers therefore aggregate data across multiple dimensions to build more robust size curves, which describe the proportional distribution of demand across sizes for a given product category, brand, gender or silhouette.

What aggregation methods do retailers use to build size curves?

Retailers typically aggregate size data along three dimensions: product hierarchy, store clustering and time period. At the product level, planners group styles by category, silhouette, brand or fit type, assuming that similar garments will share similar size distributions. For example, slim-fit trousers and regular-fit trousers may exhibit different size curves, so separating them improves forecast accuracy.

Store clustering groups locations with similar demographic and size demand profiles. A flagship store in an urban centre may see stronger demand for smaller sizes, while suburban or regional stores may skew towards larger sizes. Clustering ensures that each store group receives an allocation that reflects local preferences rather than a single global average.

Time aggregation smooths seasonal and promotional effects. Planners may pool data from multiple equivalent periods, such as the past three spring seasons, to build a representative baseline curve. This reduces the impact of one-off events, such as a viral social media post or weather anomaly, that distort a single season's data.

How do curve smoothing and machine learning refine size forecasts?

Once aggregated, raw size distributions often remain noisy. Curve smoothing algorithms apply statistical techniques such as moving averages, exponential smoothing or spline fitting to produce a continuous, stable size curve. These methods assume that true demand follows a relatively smooth distribution, such as a bell curve or skewed normal distribution, and that sharp spikes or troughs in the data reflect measurement error or stock-outs rather than genuine preference.

Modern forecasting platforms, including those from Blue Yonder and Relex, embed machine learning models that learn from actual sell-through and stock-out patterns. These systems can adjust size curves dynamically during the season, detecting when certain sizes are selling faster than expected and recommending reorders or reallocations. By incorporating signals such as sell-through velocity, inventory levels and markdown activity, the models refine their estimates of unconstrained demand, the quantity customers would have purchased if stock had been unlimited.

What impact does better size forecasting have on merchandising performance?

Improved size-level forecasting directly affects two key merchandising metrics: markdown rate and full-price sell-through. When retailers allocate inventory that matches local demand, they reduce the need for end-of-season clearance sales and the lost margin that accompanies them. Industry practitioners report that refining size mix can reduce markdowns by 5 to 15 per cent, though the exact benefit depends on category, channel and the baseline accuracy of existing allocation processes.

Conversely, better size forecasting mitigates stock-outs of high-demand sizes, which translates into higher full-price capture and improved customer satisfaction. A stock-out not only loses the immediate sale but may also drive the customer to a competitor, eroding loyalty over time. By ensuring that popular sizes remain in stock throughout the selling period, retailers protect revenue and market share.

How do pre-season and in-season forecasting differ?

Pre-season forecasting occurs during assortment planning and initial allocation, when merchandisers must commit to size quantities before any actual sales data is available. At this stage, forecasts rely entirely on historical data, product attributes and market research. Planners use aggregated size curves from comparable past styles to estimate demand for new introductions, adjusting for anticipated trends such as changes in customer demographics or shifts in fit preferences.

In-season forecasting updates predictions as real sales data accumulates. After the first few weeks, merchandisers can observe which sizes are selling quickly and which are lagging, then adjust replenishment orders or reallocate stock between stores. Advanced planning systems automate this process, continuously recalculating size curves and generating recommendations for inventory moves. This adaptive approach reduces the risk of both excess inventory and missed sales, especially for core or replenishable items with longer selling windows.

The table below summarises key differences between pre-season and in-season size forecasting:

What are the practical challenges in implementing size forecasting?

Despite the benefits, several operational challenges complicate size forecasting. Data quality remains a persistent issue: inconsistent size labelling across brands, missing stock-out flags and incomplete transaction records all degrade model performance. Retailers must invest in data governance to ensure that forecasting systems receive clean, reliable inputs.

Organisational alignment also matters. Merchandisers, planners and allocation teams may have conflicting objectives or use different performance metrics, leading to friction when models recommend changes to established workflows. Successful implementations require cross-functional collaboration and a shared commitment to testing and learning from forecast accuracy over multiple seasons.

Finally, model complexity and interpretability trade off against one another. While sophisticated machine learning algorithms may deliver incremental accuracy gains, they can also become black boxes that planners distrust. Many retailers favour transparent, rules-based approaches or hybrid models that combine statistical methods with human judgment, ensuring that planners retain control and can explain allocation decisions to store managers and suppliers.