Skip to content

Case Studies Retail

Demand forecasting that cut stockouts across 200+ stores

We built a forecasting platform on Databricks that combined POS, weather, and promotional data into a single demand signal.

22% fewer stockoutsForecast refresh time cut from days to hoursUnified data model across 200+ stores

Client challenge

What was wrong with the legacy environment

Store-level demand forecasts relied on spreadsheets and static rules

POS, promotional, and weather data lived in separate systems with no shared model

Forecast refreshes took days, too slow to react to demand shifts

No visibility into forecast accuracy by store or category

Modernization approach

How Karsient approached the transformation

Karsient built a unified demand-signal pipeline that blends POS transactions, promotional calendars, and external weather data into a single feature set, replacing spreadsheet-driven planning with a governed, automated model.

Architecture

The modern target architecture

1

Batch and near-real-time ingestion from POS and inventory systems

2

Bronze/Silver/Gold medallion layers for raw, cleaned, and modelled demand data

3

Forecasting models orchestrated on a daily schedule via Databricks Workflows

4

Power BI dashboards for store- and category-level forecast accuracy

Migration

Migration & re-engineering strategy

Legacy spreadsheet workflows were replaced store-cluster by store-cluster, with each cluster's new forecast validated against the prior manual process for two full cycles before cutover.

Engineering improvements

What changed under the hood

Automated daily forecast refresh replacing manual weekly cycles

A single demand-signal feature store shared across planning teams

Parallel processing across store clusters cutting refresh time

Forecast-accuracy monitoring built into the pipeline itself

Business impact

Measurable outcomes

22% reduction in stockouts across 200+ stores

Forecast refresh time cut from days to hours

One consistent demand model replacing per-region spreadsheets

Planners now spend time on exceptions, not data assembly

Technology

Technology used

Databricks
Apache Spark
Delta Lake
Power BI
Python

Want results like these?

Let's scope your next data or AI initiative