Solution · Stock planning
Stock levels based on real demand, not guesswork.
The forecasting model uses sales history, seasonality and available capacity to recommend quantities for each item and location. Forecasts cover horizons of 4, 12, 26 and 52 weeks and flag looming shortages or surpluses. In the project we delivered, the client's figures show revenue growth of 5–6%.
Typical complexity level Level 3 · Complex · from 6,490 EUR excl. VAT
When it makes sense
- Some items are out of stock when needed, while others tie up money unnecessarily
- Orders are planned manually based on experience
- You have multiple warehouses, branches or points of sale
- Seasonality and promotions throw off manual estimates
How it works
- 01
We start from sales and stock history.
Item movements, stock levels and minimum quantities.
- 02
We account for seasonality and other factors.
Time series with seasonality and other factors.
- 03
The model forecasts demand weeks ahead.
Models for 4-, 12-, 26- and 52-week horizons.
- 04
You get recommended order quantities.
Quantities and alerts, exported to the ordering system.
What we build on
Python, time series models, scikit-learn and LightGBM, a data warehouse and PostgreSQL. We work with aggregated data without any customer personal data.
In practice · Inventory planning
Stock based on real demand
The forecast uses historical data, seasonality and available capacity. The team gets a recommended quantity for every item and location.
Feasibility study completed
Starting point
Some ordered items were not available in time, others tied up money unnecessarily. Manual estimates could not reflect all differences between items and locations.
Goal: Improve item availability while reducing unnecessary stock.
Solution
The model forecasts demand for individual items and locations. It links the recommendation with stock levels, available capacity and minimum quantities.
Results
- The team received a recommended quantity for every item and location.
- The forecast covers four time horizons.
- The project passed a feasibility study on historical data.
In the project, the documentation reports a 5–6 % increase in revenue.
Technology
- Python and time-series models
- scikit-learn / LightGBM
- Data warehouse and PostgreSQL
- Dashboard and export to the ordering system
Security and data
- Aggregated data without customers' personal data
- Runs in the customer's infrastructure
- Accuracy validated on historical data
Operations
- Design for regular retraining and batch calculations
- Accuracy evaluation by time horizon
- Option to extend to further item groups
This case study is anonymised. We do not name the company or its industry.
Frequently asked questions
How much history do we need?
Ideally at least two years, so the model can capture seasonality. With a shorter history, you can start with simpler models.
How often is the forecast recalculated?
Usually once a week in a batch. We retrain the model regularly and monitor its accuracy.
You might also be interested in
Solution · Data evaluationData analysis and quality control
Machine learning models compare measured data and help tell measurement errors from real deviations. A dashboard for expert review, run on your infrastructure.
Level 3 · from 6,490 EURMore
Solution · Scenario comparisonProcess and capacity simulation
Simulate workstation layouts, capacity and utilisation in the browser. Compare scenarios before changing real operations. Discrete-event simulation, 3D view.
Level 4 · from 12,900 EURMoreNot sure where to start? That's fine.
Get in touch. In a short call, we'll find out what takes up most of your time today and suggest a first step. The consultation is free and with no obligation.