ERP + POS
daily
- feeds
- Feature build
All case studies
ForecastingOptimizationAnalytics4 months
Demand Forecasting + Inventory Optimization
Planning ran on a spreadsheet and one growth rate. Stock now follows actual demand.
Weighted MAPE
2.7%
Reconciled hierarchical forecast
Was
12.1%
Spreadsheet with blended growth rate
In plain terms
What this actually was.
No jargon in this section. The technical write-up is further down.
What was going wrong
Planning used spreadsheets and a single growth assumption applied to everything. Popular lines ran out while slow ones sat in the warehouse tying up cash, and nobody could put a number on the trade-off between the two.
What we built
We built forecasting that works product by product and adds up consistently to category and company totals, plugged into the planning process people already used, and able to answer what-if questions.
What changed
Forecast error fell from 12.1% to 2.7%. Cash previously tied up in the wrong stock was released without more items running out, and planners now review the exceptions instead of rebuilding spreadsheets.
Architecture
How it fits together.
Scroll through the stages. Anything marked as added is a component that did not exist before this project.
Feature build
calendar + promo
- receives from
- ERP + POS
- feeds
- Forecast model
Forecast model
hierarchical
- receives from
- Feature build
- feeds
- Reconciliationper level
Reconciliation
coherent
- receives from
- Forecast modelper level
- feeds
- Stock optimiser
Stock optimiser
service level
- receives from
- Reconciliation
- feeds
- Planner UI
Planner UI
exceptions
- receives from
- Stock optimiser
ERP + POS
daily
- feeds
- Feature build
Feature build
calendar + promo
- receives from
- ERP + POS
- feeds
- Forecast model
Forecast model
hierarchical
We added this- receives from
- Feature build
- feeds
- Reconciliationper level
Reconciliation
coherent
We added this- receives from
- Forecast modelper level
- feeds
- Stock optimiser
Stock optimiser
service level
We added this- receives from
- Reconciliation
- feeds
- Planner UI
Planner UI
exceptions
- receives from
- Stock optimiser
The write-up
How it was actually built.
Dataset, approach, measured results and the stack. Written for whoever has to review it.
ForecastingOptimizationAnalytics4 months
Demand Forecasting + Inventory Optimization
Forecasting that plugs into planning workflows and supports what-if scenarios.
Problem
Planning ran on spreadsheets with a single blended growth rate. Fast movers stocked out while slow movers tied up working capital, and nobody could quantify the tradeoff.
Dataset
- rows
- 31M SKU-days
- features
- 178
- sources
- ERP, POS, promotions, weather
- window
- 48 months
Approach
- 01Built a hierarchical forecast reconciled across SKU, store and region
- 02Backtested with rolling-origin evaluation rather than a single split
- 03Modelled promotions and calendar effects as explicit regressors
- 04Fed forecasts into a safety-stock optimiser with service-level targets
Metrics
- 2.7%
- weighted MAPE -9.4pp
- 97.3%
- forecast accuracy 12-week horizon
- -34%
- stockouts on A items
- -21%
- inventory held same service level
Business impact
- Working capital released without hurting availability
- Planners review exceptions instead of rebuilding spreadsheets
- Promotion decisions backed by a measured lift estimate
Tech stack
- Python
- Prophet
- CatBoost
- dbt
- BigQuery
- Looker
- Airflow
Next
Need a system like this?
Answer six questions and we will tell you whether this shape fits your problem — including when it does not.