Case Study · FMCG
Hayat Kimya: weekly e-commerce order forecasting on Microsoft Fabric
A system that forecasts marketplace and e-commerce orders weekly, at the granularity the warehouse shift plan needs, and runs end to end as a single Fabric pipeline from data warehouse to report.
- Weekly
- Forecast horizon
- One pipeline
- Automation
- Every week
- Model tracking
Campaign calendar as a model input
From data pull to report, no human intervention
Actuals next to the forecast
Context
Hayat Kimya sells bundled products through several marketplaces and e-commerce sites. These orders are picked by warehouse staff at the logistics warehouse, and daily staffing need depends directly on the number of incoming orders. Order volume, in turn, swings sharply from week to week with marketplace campaigns.
Problem
When order volume is unknown in advance, shift planning fails in both directions: too few people delay shipments, too many raise cost. What was needed was a system the planning team could trust, producing a weekly order forecast in the operation’s own breakdown.
Solution
The system was built end to end on Microsoft Fabric; no step from data warehouse to report leaves the platform.
Data source. A medallion-architecture (Bronze / Silver / Gold) data warehouse on Fabric. Sales and order data are taken from here.
Data preparation. Data cleaning, analysis and feature engineering in notebooks. Because campaigns are the main driver of demand, the campaign calendar was taken from Excel and added to the model as a feature — this step was decisive for the model’s accuracy.
Forecast model. Machine-learning forecasts are produced weekly at the hierarchy the planning team needs and written to the Lakehouse. Not a company-wide total, but a table at the level the shift plan is made.
Model tracking. A Power BI report puts actual sales next to the forecast every week; when the model starts to drift, it shows in the first week.
Automation. Data pull, preparation, forecasting and reporting were built as a single Fabric pipeline and scheduled weekly. The process runs without human intervention.
Outcome
Warehouse staffing is now planned on the weekly forecast output. The forecast became a tool the planning team uses directly, on one platform from data warehouse to report.
Because model accuracy is monitored continuously, the system works as a maintainable operational component rather than a forecast that decays over time.