Demand forecasting and distribution optimization
Distribution and logistics, multiple markets
Michael Jan Rogocki delivered this project before founding cm-opti. We use the same method in client projects today.
Industry and type of organization
A company distributing goods across many markets at once. It plans inventory under fluctuating demand, where too little stock means gaps on the shelf and too much locks up capital in the warehouse.
Business problem
Planning inventory for many markets at once is hard, because each market has its own seasonality, its own sales pace and its own supply constraints. With manual planning the decisions rest on averages and on the planner's experience, and those do not keep up with the variability.
The result shows at both ends of the warehouse: where the goods ran out and the sale was lost, and where stock sits that no one will take this season. Both cost money.
Approach: demand forecasting and supply optimization
The project combined two methods. The demand forecast was based on classic machine learning, mainly on regression models that predict a numeric value, in this case the expected sales of a single article. The model takes into account sales history, seasonal patterns and trends over time (time series analysis) as well as external factors that influence demand. On top of that forecast, mathematical optimization algorithms plan the distribution of goods, taking into account logistics costs, lead times and warehouse capacity.
This is stochastic optimization: it takes into account that the forecast is uncertain and plans with that margin in mind. The planner gets several scenarios for how demand and inventory could develop and sees what happens under different assumptions before making a decision.
Impact: optimizing warehouse and logistics processes
article level
forecast for a single product
instead of averages for whole product groups
multiple markets
one forecasting approach
each market with its own seasonality
A better forecast means less stock sitting in the warehouse and deliveries planned for actual demand rather than averages. Products move before they have to be marked down, so less money is lost on surplus and clearance sales. Stock turns over instead of getting stuck, which translates into a higher margin.
Scale of the implementation
The forecast covers the assortment at the article level and many markets at once. The models need retraining as new sales data comes in, because a forecast based on outdated patterns quickly loses accuracy.
If you plan inventory for many markets and end up guessing between stockouts and overstock, describe your case to us. We will check whether your sales data is enough to forecast more accurately.
Related articles in the cm-opti knowledge base
Service area: Machine learning and forecasting (data analysis, forecasting and process optimization).