FMCG case study

Category Management Analytics for Lorenz Polska: ML Forecasts That Beat Manual Estimates

Lorenz Polska, a B2B snack company serving major retail chains, needed category management analytics to automate assortment preparation, understand shopper behavior and grow gross margin. DS Stream built a modular analytics ecosystem on Google Cloud Platform. Machine learning forecasts consistently outperformed manual sales estimates, and Lorenz's team now develops the platform on its own.

Discuss your category data See more case studies

Results at a glance

Better demand planning: machine learning models consistently outperformed manual sales estimates.
Richer promotion data: LLMs unlocked value from promotion datasets that had been siloed.
Faster market response: automated assortment change detection is deployed in selected retail chain stores.
Lasting in-house skills: Lorenz's team now develops and adapts platform features independently.

What was Lorenz Polska trying to achieve in category management?

Lorenz Polska sells snacks B2B to major retail chains, with modern trade as its priority channel. The company wanted strong analytics inside its commercial teams, automated assortment preparation for different retailers, a deep view of consumer purchasing, higher gross margin across the portfolio, lower assortment management costs and a reliable way to measure promotion effectiveness.

Constraints that shaped the project

Organization: analytics expertise was still growing inside the company, and teams were running several initiatives in parallel.
Data: inputs came in different granularity and formats, so datasets needed harmonizing before SKU-level and promotional insights could be enriched.
Scale: data volumes were in gigabytes, not terabytes. Tailored, efficient tools fit better than heavy big-data infrastructure.

What did DS Stream build for Lorenz Polska?

DS Stream designed a modular, scalable and pragmatic analytics ecosystem on Google Cloud Platform (GCP). It avoids unnecessary complexity but leaves room to grow. Each module answers one category question, from who buys what to which SKUs belong in a retailer's assortment.

Core components

Customer and product segmentation with K-Means, DBSCAN and hierarchical clustering.
SKU-level demand forecasting that accounts for seasonality and region, using ARIMA and LightGBM.
Price elasticity modeling with a Bayesian GLM.
Substitute and complement detection through correlation and semantic analysis.
SKU portfolio optimization with linear and quadratic programming.
Automated data integration (ETL in BigQuery) feeding interactive dashboards for business users.

How was the platform delivered and handed over to Lorenz's team?

Delivery was built on knowledge transfer and co-creation. Every analytical module came with hands-on workshops, so Lorenz's team could enhance and expand the platform without outside help. Work ran in agile, iterative cycles, and each prototype was designed to scale and extend.

Planned extensions

The prototypes leave room for future use cases named in the project:

Shelf analytics
Integration with Nielsen data
Advanced pricing strategies

Which technologies power the platform?

The stack was chosen to match Lorenz's gigabyte-scale data, which kept costs down without limiting results.

Google Cloud Platform as the cloud foundation
BigQuery / BigQuery ML for data integration (ETL) and in-warehouse modeling
Python with pandas, scikit-learn, LightGBM and PuLP for modeling and optimization
Vertex AI and Jupyter / Vertex AI Notebooks for model development
SQL for data preparation

Services involved: Data Engineering, Machine Learning, Data Science and Cloud Solutions.

FAQ about the Lorenz Polska project

Why didn't Lorenz Polska need big-data infrastructure?

Lorenz Polska works with data in gigabytes, not terabytes. At that scale, heavy big-data infrastructure would add cost and complexity without adding value. DS Stream chose a pragmatic stack on Google Cloud Platform, built around BigQuery, Python and Vertex AI, that stays cost-effective and can still scale later.

What role did LLMs play in the project?

LLMs were used for semantic enrichment of existing data, especially promotion datasets that had been siloed. This enrichment also supports the detection of product substitutes and complements. It let Lorenz Polska get more value from data it already had, without new data collection.

Can Lorenz Polska extend the platform on its own?

Yes. Each module was paired with hands-on workshops, and Lorenz's team now develops and adapts platform features independently. The modular design was built with future use cases in mind, such as shelf analytics, Nielsen data integration and advanced pricing strategies.