Retail analytics platform for a Japanese supermarket chain

A Japanese supermarket chain with 420 stores generated enormous volumes of sales data, yet most replenishment decisions still depended on each store manager's experience. Reports were compiled by hand in spreadsheets and typically reached the executive team a week late. The company partnered with Zentrova's Tokyo centre to turn its data into a tool for everyday decisions.

The challenge

Data lived in 38 separate sources: point of sale, loyalty cards, suppliers, logistics and even weather feeds. Fresh produce accounted for 35% of revenue but suffered high waste, while core items regularly ran out on weekend shelves. Every department had its own definition of metrics such as gross margin or sales per square metre, so meetings started with arguments about the numbers. Head office wanted a single version of the truth, delivered early enough each morning to change what was ordered that day.

Our solution

• Data lakehouse architecture in the cloud, ingesting all 38 sources at intervals ranging from 15 minutes to daily.
• Shared semantic layer built with dbt, standardising more than 160 business metrics with automated data quality tests.
• Demand forecasting models in Python for every item in every store, accounting for seasonality, promotions and weather.
• Role-based dashboards for executives, regional managers and store managers, designed to work well on tablets.
• Access control by store and region, with loyalty member data encrypted in line with personal data protection rules.
• Phased rollout: a six-week pilot in 20 stores, measured carefully before extending to the whole chain.

Results

The platform was in use across the entire chain after seven months, with 2,300 internal users logging in every week.

• Average inventory value reduced by 18%.
• Fresh produce waste down 27%.
• Shelf out-of-stocks on core items down 34%.
• Executive reports ready at 7 a.m. every day instead of a week late.
• Store-level demand forecast accuracy of 91%.

The forecasting models are retrained every week and their accuracy is tracked openly on the same dashboards, so store managers can see when to trust a suggestion and when to override it. The client's analysts now build new reports themselves on the shared semantic layer, and the next phase extends forecasting to supplier ordering and promotional planning, with Zentrova's Tokyo team on hand for model tuning.

Our store managers now start the day with data instead of guesswork. — Chief Operating Officer, client supermarket chain