Size Curve Optimization

The Size Curve Problem

Traditional retail approaches treat size distribution as uniform—send every store the same percentage of XS, S, M, L, and XL inventory. This one-size-fits-all methodology seems logical when based on national population statistics, but it ignores a fundamental reality: customer demographics and preferences vary dramatically by location, channel, and product type.

The result? Predictable, costly patterns emerge across retail networks. College town stores perpetually stock out of small sizes while drowning in excess large inventory. Suburban family locations run short on medium and large sizes. Urban flagship stores serve different customer segments than outlet centers, yet receive identical size allocations. These systematic mismatches translate directly into lost sales, excessive markdowns, frustrated customers, and expensive inter-store transfers.

Store-Specific Size Demand

Each store has a unique customer base with distinct size preferences reflecting local demographics, customer segmentation, and shopping patterns. A downtown professional store attracts different body types and style preferences than a college campus location or a suburban family shopping center.

Understanding Local Demand Drivers

Demographics matter. Age, income, ethnicity, and lifestyle create measurable differences in size distributions. Younger populations skew smaller. Areas with specific ethnic concentrations show different average body types. Higher-income neighborhoods often correlate with more fitness-focused lifestyles and corresponding size preferences.

Store format influences selection. Flagship locations in urban centers serve style-conscious customers seeking the latest trends in fitted silhouettes. Outlet centers attract value shoppers less concerned with cutting-edge styles, often showing broader size distributions including more extended sizes. Mall stores in family-oriented suburbs see balanced middle-size demand.

Competitive context shapes traffic. A store near a specialty plus-size retailer may see different size patterns as some customers segment themselves to competitors. Conversely, being the only option in a geographic area creates more representative demand across all sizes.

Machine learning models identify store-specific size curves from historical sales data while controlling for stockout bias—the phenomenon where depleted sizes appear to have lower demand simply because they weren't available. By estimating unconstrained demand rather than merely observed sales, these models reveal true customer preferences.

Product-Specific Size Curves

Size demand varies not just by store, but by product type—even within the same location. The size distribution for athletic wear differs fundamentally from business casual or evening wear. Product characteristics drive these variations in predictable ways that machine learning can capture.

Category-Level Variations

Athletic and activewear typically shows a flatter, smaller-skewed distribution. Compression fabrics and performance fits encourage size-down behavior. Customers purchasing workout clothing often aspire to or maintain slimmer body types, shifting demand toward XS and S.

Outerwear shifts larger because layering requires additional room. A customer who wears Medium tops often purchases Large jackets. Cold weather markets show more pronounced shifting as heavier layering is common. This category also sees longer purchase cycles, with customers holding inventory across seasons.

Denim presents unique complexity. Two-dimensional sizing (waist and inseam) creates a matrix rather than a simple curve. A 32-inch waist might need a 30-inch inseam in one store but a 32-inch inseam in another, depending on customer height distributions. Stretch versus rigid denim further complicates the picture, as stretch enables size-down behavior impossible with traditional fabrics.

Women's categories show higher variability than men's due to less standardized sizing, more frequent style changes, and broader acceptance of size fluidity. The same customer might wear Small in one brand's tops, Medium in another, creating noise that machine learning must filter to identify true demand.

The models learn these product-specific patterns by analyzing size sell-through rates across categories, brands, and styles. This enables accurate curves even for new products without sales history—the system predicts based on similar products' historical performance, adjusted for the specific store's customer base.

Handling Size Stockouts

A critical challenge in size optimization is the stockout bias problem. Historical sales data shows what sold, not what customers wanted. When Small sells out in week two of a twelve-week season, you've captured only a fraction of potential Small demand. The sales data falsely suggests lower Small demand compared to sizes that remained in stock.

The Censored Demand Challenge

Standard analytical approaches treat zero sales as zero demand. But zero sales when inventory exists means genuine lack of interest, while zero sales after stockout means suppressed demand. Conflating these scenarios produces systematically biased size curves that perpetuate the original allocation mistakes.

Machine learning models correct for this censoring by modeling latent demand rather than just observed sales. They identify when sizes likely stocked out by examining inventory trajectories and sell-through velocity. If Small sold its last unit in week two while Medium still had inventory in week twelve, the model infers substantial unmet Small demand.

This correction is particularly important for new product launches where initial allocations are most uncertain. Early stockouts provide valuable demand signals that must be properly interpreted to improve subsequent replenishment and future product allocations.

Cross-Store Size Reallocation

Despite optimal initial allocation, size imbalances inevitably develop as actual sales patterns deviate from forecasts. Some variation is random noise, but persistent patterns indicate opportunities for beneficial transfers. Machine learning enables dynamic reallocation, transforming size inventory from static to fluid, flowing toward demand continuously.

When to Transfer

Not every size imbalance warrants a transfer. Transfers incur real costs—handling, shipping, timing delays, and disrupted presentation. The expected benefit must exceed these costs. Machine learning models identify beneficial transfers by analyzing current inventory positions, predicted remaining demand, and transfer costs.

The optimization considers the full network simultaneously. Rather than bilateral store-to-store decisions, it solves for the best set of transfers across all locations to maximize expected profit. This might mean Store A sends Medium to Store B, Store B sends Large to Store C, and Store C sends Small to Store A—a complex web impossible to identify manually.

Dynamic Reallocation in Practice

Weekly evaluation cycles assess transfer opportunities as new sales data arrives. Early weeks see more transfers as initial allocation errors become apparent. Mid-season transfers capture emerging opportunities from unexpected demand patterns. Late-season transfers focus on consolidating remaining inventory to strongest locations.

Automated execution is essential at scale. A national retailer might evaluate thousands of potential transfers weekly. Manual review is infeasible. Machine learning systems automatically flag high-value transfers while allowing human oversight for unusual situations or high-stakes decisions.

Performance tracking measures transfer effectiveness. Did the transferred size sell through at the destination? Would it have sold at the origin? What was the realized margin impact? These metrics train the system to improve transfer decisions over time.

Measuring Size Optimization Success

Effective size optimization should deliver measurable improvements across multiple dimensions. Track these metrics to evaluate program success and identify opportunities for refinement.

Key Performance Indicators

Implementation Considerations

Data Requirements

Successful size optimization requires clean, comprehensive data. At minimum, you need historical sales by size, store, product, and time period. Enhanced results come from inventory positions, customer returns by size, and store attributes (demographics, format, square footage). The more data, the better the models, but start with what you have and expand over time.

Organizational Readiness

Size optimization touches multiple functions—merchants, planners, allocators, store operations, and supply chain. Success requires cross-functional alignment on goals, processes, and decision rights. Resistance often comes from merchants comfortable with traditional intuition-based allocation. Demonstrate results through controlled pilots before full-scale rollout.

Technology Stack

Enterprise allocation systems, data warehouses, and machine learning platforms form the technology foundation. Cloud-based solutions offer flexibility and scalability without heavy upfront investment. Many retailers start with their existing allocation systems enhanced with size optimization modules, then evolve to more sophisticated platforms as capabilities mature.