Smart Merchandising with ML: Optimize Product Mix for Every Store

Quick Summary

Challenge

Retailers were missing revenue due to poor stocking decisions. Some stores ran out of promoted products, others stocked what didn’t sell

Solution

Tatras Data built an ML model that continuously recommends the optimal product mix per store, based on local demand, sales velocity, and demographic patterns

Result

23% reduction in out-of-stock promoted products.

Tech Stack

AI: Neural networks Product-similarity models Demand forecasting | ML: Sales-to-SKU ratio estimation | Data & Retrieval: Store-level sales data Online conversion logs Inventory history | Dev: Python pandas Scikit-learn XGBoost | Viz: Store dashboards for merchandisers and planners | Deployment: Integrated into internal stocking recommendation engine

The Challenge

Product mix is one of the most overlooked drivers of store profitability. A U.S.-based retailer wanted to optimize what to stock, where, and when. Some stores were running out of high-demand products. Others stocked items that sat untouched. Regional preferences weren’t reflected in stocking decisions, and new product introductions were largely guesswork. The results were entirely predictable: Missed sales, frustrated customers, and wasted shelf space.

A Day in the Life: Before Our Solution

Merchandising teams reviewed inventory with limited visibility. Some stores ran out of high-demand products, while others were overstocked with slow-moving items. Stocking decisions relied on historical data and intuition, with little alignment to local demand. Promotions and new product launches often missed the mark, leading to lost sales and inefficient shelf space.

Pain Points:

Out-of-stock on promoted items during peak demand
Overstocking of low-demand products consuming shelf space
Lack of real-time adaptation to shifting regional preferences
New product stocking based on guesswork, not data
Lost sales due to mismatch between customer intent and store availability

Solution

1. Core Innovation

Tatras developed an ML engine that dynamically forecasts product-store fit:

Built SKU-level demand forecasting model based on store sales + online behavior.
Identified underperforming SKUs relative to local demographic fit.
Created proxy signals for new product potential using similarity to past performers.
System continuously updates based on actual sell-through and local shifts.
Output: A per-store stocking recommendation that optimizes inventory for margin and turnover.

2. Key Features

SKU-to-store mapping based on local sales potential.
Predictive analytics for new product success.
Continuous optimization based on observed sell-through vs forecast.
Demographic proxy modeling using online purchase data.
Planner dashboard for review, override, and insight

3. Workflow Integration

Each week, the model generates a stocking recommendation per store. Planners can review changes, compare to last week’s shelf mix, and approve with a click. Store managers now receive SKUs that are far more aligned with what their local customers are looking for.

Outcomes

✅ 23% decrease in out-of-stock incidents for promoted products

📈 Improved sales performance by region

🧠 Smarter launches for new products based on proxy demand