Pricing the Whole Store — Demand-Transfer Forecasting at Fortune 100 Scale

Challenge

A Fortune 100 retailer brought me in to tackle a complex business problem with significant revenue implications as they furthered their forays into data science. The challenge: building a dynamic pricing model that could predict how price changes would impact not just individual product sales, but the entire ecosystem of related products:

• Similarly priced substitute items
• Other products from the same vendor across price points
• Related items in other departments

This complex demand transfer could flow between products, vendors, the client's own stores, or even to competitor retailers. Each effect could trigger cascading changes across thousands of SKUs in up to 2,000 stores, forecasted for 13 fiscal weeks.

Strategic Approach

The initial hand-crafted model did not provide acceptable performance for web users due to the size and scope of the predictor, and so we needed to combine technical expertise with user-focused interaction design:

• Resource Optimization: We made low-level changes to how data was stored, and added a batching strategy to more efficiently load the data into GPU memory, and designed intelligent priority queues to ensure large requests did not block smaller ones.

• Frictionless Integration: We implemented a websocket API enabling users to initiate multiple complex simulations and asynchronously receive results without blocking their workflow. This allowed merchandisers to compare multiple scenarios, while dramatically improving both system scalability and user experience.

• MLOps Architecture: We established robust cloud-based pipelines for ETL of source data, training, validation, and seamless model deployment, so engineers could spend more time on what mattered.

Business Impact

This initiative delivered results across multiple dimensions:

• Reduced forecasting experiment time from hours to minutes for merchandising associates
• Reduced model update cycles from weeks to hours for improved A/B experimentation
• Created significant operational efficiencies by empowering business users with self-service capabilities
• Enabled automated anomaly detection for price changes that would typically escape human oversight
• Allowed deeper and more varied regional market-based strategies
• Future potential to allow agentic AI to decide which of these experiments and strategies to pursue

Culture Shift

As an early adopter in the organization, this team and partners in our infrastructure group were able to establish best practices in team design, cloud spending, and even a command-line SDK that provided a "paved road" for future ML initiatives at the company. By reducing users' barriers to using the tool, we decreased our lead time in delivering features and increased model reliability.

This project exemplifies my approach to engineering leadership: a novel use of existing data, that empowers users to improve business outcomes, while trailblazing engineering excellence across the organization.