Smarter Size Curves for Retail Inventory Optimization

If you’ve ever been frustrated by how size curves misrepresent actual demand, you’re not alone.

Most merchandising systems calculate size curves directly from sales data. That sounds logical, until you realize how often sales don’t reflect intent. If a certain size is consistently out of stock, it’s not just underrepresented in your data — it’s invisible. And that invisibility leads to missed demand, misallocated inventory, and recurring stockouts, hindering effective retail inventory optimization.

Imagine a T-shirt that appears to sell 20% small, 60% medium, and 20% large. But large has been sold out half the time. That distribution isn’t demand, it’s a data artifact. The actual appetite for large might be closer to 40%. And yet most systems keep using the distorted sales ratio to make future decisions, reinforcing the very problems they were meant to avoid.

This is the trap of article-level size curves. At Syrup, we wanted to build something better. I’m excited to introduce you to our Bulk Updates feature, powered by smart size curves.

Why Traditional Size Curves Fall Short

When you rely entirely on sales data to define size curves, you inherit its flaws. Stockouts bias the curve. Low-volume articles produce noisy, unreliable ratios. And even within the same size set, different articles behave very differently. One S/M/L style might sell 50% in medium, another might peak in large, but traditional systems treat them the same.

Worse, if you use actual sales to build size curves and then use those same curves to forecast future demand, you’re leaking the target into the input. It’s a feedback loop that can overfit and mislead.

We knew we needed a new approach — one that could generalize across articles, account for inventory constraints, and adapt to changes in available sizes. That’s what led us to develop a smarter, model-based size curve.

A Smarter Size Curve: Embeddings Meet Merchandising

At the heart of our solution is a machine learning model that generates size curves using article similarity and historical patterns, not just raw sales ratios. It starts by embedding both visual and textual attributes of a product: things like category, color, silhouette, and description. These embeddings are trained in-house to capture meaningful relationships between items.

Given a new article, the model looks for similar items in this embedding space and learns from their historical size distributions. Crucially, it only considers available sizes, so if the product is offered in M, L, and XL, it predicts a curve within that scope. There’s no hallucinating demand for sizes that aren’t in the assortment.

Because the model learns from similar articles, it doesn’t need extensive sales history for the specific SKU. And when there is little to no data — such as in a new store — it falls back to more general curves built at the category or segment level. This means you still get an informed size distribution even in cold-start scenarios.

It’s a flexible, resilient system that avoids the pitfalls of article-level definitions while staying grounded in actual selling behavior.

Turning Insights into Action: Bulk Updates

Building better size curves was the first step. The next was helping teams use them — quickly, easily, and in the flow of work.

That’s where our new Bulk Updates feature comes in.

With Bulk Updates, merchandisers can adjust the total quantity of any article at any store and let the system handle the intelligent distribution of units across sizes. Instead of manually tweaking size-level allocations, you simply enter a new total, and the model applies the appropriate size curve in real time.

These updates apply across our Initial Allocation and Replenishment workflows — and they’re available directly in the same table where allocators already make adjustments. No extra steps. No switching tabs. Just smarter, faster updates powered by better data.

Key features include:

• Article-Level Quantity Adjustments — Edits the total recommended units for a product at a specific store with a single input
• Smart Size Distribution Algorithm — Uses historical selling patterns and our “Greedy Rounding with Fairness” method to maintain proportional, accurate size representation
• Hierarchical Curve Logic — Automatically applies the most relevant size curve available: article-store, category-store, or overall category
• Integrated Workflow Support — Works seamlessly across allocation, replenishment, and rebalancing use cases

How It Works: Real-World Retail Inventory Optimization Scenarios

In one case, a merchant is increasing inventory for a top-performing store. They raise the article quantity from 108 to 150 units. Rather than requiring manual edits across each size, the system automatically scales up the size distribution based on that store’s historical pattern, preserving the shape of the curve while adjusting depth.

Summary:

• Scenario: Boosting allocation for a strong store
• Action: Increase from 108 to 150 units
• Result: Units are proportionally scaled across sizes using store-specific history

In another case, they’re reducing allocation for a lower-performing location. They drop the total from 13 to 8 units. The model responds by concentrating those units in the core selling sizes, potentially zeroing out fringe sizes like XS or XXL, depending on that store’s history.

Summary:

• Scenario: Decreasing inventory for a low-performing store
• Action: Reduce from 13 to 8 units
• Result: Concentrates units in high-performing sizes; trims fringe sizes

Even in edge cases, the system performs reliably. For example, if a newly opened store has no size sales history, the model taps into segment-level patterns from similar locations. You get a reasonable, data-driven curve — without needing to start from scratch.

Summary:

• Scenario: New store with no sales history
• Action: Enter desired total units
• Result: Model applies size curve based on segment-level patterns

And for small quantities where perfect distribution isn’t possible, our “Greedy Rounding with Fairness” algorithm ensures that allocations remain proportional. A 3-unit order might not divide evenly, but the algorithm intelligently prioritizes the sizes with the highest fractional remainder, so every unit still contributes to a representative curve.

Why This Matters — And What’s Next

When your size curves reflect actual demand — and your allocation tools are built to use them — you unlock better decisions across the board.

You can forecast more accurately. You can move faster, with confidence. You can avoid the endless whack-a-mole of adjusting individual sizes after the fact. And you can trust that your systems are accounting for realities like stockouts, sparse data, and new stores, not just replicating historical sales.

This release is part of a broader vision at Syrup: building tools that help merchandising teams work with speed and intelligence for advanced retail inventory optimization. We’re embedding machine learning directly into everyday workflows — not to replace decision-makers, but to support them with sharper insights and better automation.

If you’re curious how smarter size curves and Bulk Updates could fit into your allocation process, we’d love to show you more.

‍