Multi-SKU demand forecasting
Forecasting at the SKU / site level over short horizons (D+1 to D+30) and long ones (W+1 to W+26). Handling of SKUs with little history or intermittent demand.
TFT · LightGBM · Prophet · SARIMA · Croston
Domain of expertise
Supply chain forecasting and optimization by ML
Demand forecasting and inventory optimization models adapted to industrial constraints: seasonality, promotions, supplier lead times, capacity constraints. From ERP ingestion to output in the replenishment system.
Capabilities
What we build
Safety stock optimization
Dynamic calculation of safety stocks and reorder points from forecasts and real supplier lead times. Automatic update in the ERP.
Constrained optimization · Monte-Carlo simulation
Early stockout detection
Alerts on stockout risks over a configurable horizon, with quantification of the probability and the criticality lead time.
Probabilistic forecasting · confidence intervals
Overstock detection
Identification of items at risk of overstock based on the forecast evolution of demand and their shelf life / obsolescence.
ABC-XYZ segmentation · depreciation models
ERP / WMS integration
Output to the SAP planning modules (MRP, MD04), Oracle, Sage, or any WMS via API. Automatic recalculation at a configurable frequency.
SAP BAPI / RFC · REST API · SFTP
Forecast error analysis
Performance dashboard per SKU and family: MAPE, RMSE, bias, trends. Identification of the hardest-to-forecast items.
Grafana · Metabase · automated report
Architecture
Reference pipeline
ERP extraction → feature engineering → model → inventory optimization → output to the replenishment system. Weekly or daily refresh depending on the horizon.
Sources
ERP / WMS / files
Sales history, stocks, open orders, real vs. nominal supplier lead times. SAP, Oracle, Sage, CSV.
Features
Exogenous variables
Seasonality, working days, promotions, prices, external data (weather, sector indices) depending on the context.
Model
Ensemble or single model
TFT for long series with covariates, LightGBM for performance on a large catalogue, Prophet for strong seasonality.
Optim.
Stock & replenishment calculation
Dynamic safety stocks, reorder points, economic order quantities under MOQ and capacity constraints.
Contexts
Problems addressed
ML delivers measurable value as soon as demand is variable, the catalogue is broad, and the historical data covers at least 12 to 18 months.
Industrial distribution: thousands of SKUs
Irregular demand, supplier MOQ constraints, long lead-time delays. ML handles the diversity of demand profiles.
Make-to-stock manufacturing: MTO / MTS trade-off
Decoupling decision per item based on demand predictability and manufacturing lead time.
Multi-site supply chain
Inter-warehouse allocation, transfers, consolidation of aggregated demand towards the production plants.
Long lead-time components
Electronics, raw materials with 8 to 20 week lead times. Long-range forecasting is critical for procurement.
Results
Observed orders of magnitude
Measured on comparable deployments. They depend on demand volatility, the quality of the historical data and the breadth of the catalogue.
20 – 40 %
reduction in stockouts over the forecast horizon
10 – 25 %
reduction in locked-up inventory on optimized items
10 – 20 %
MAPE over a 4-week horizon depending on demand volatility
2 – 4 weeks
between project kick-off and the first testable forecasts
Approach
How we proceed
01
Data audit
Quality and coverage of the historical data, inventory of available exogenous variables, identification of priority SKUs. 1 week.
02
Baseline
Reference model (naive or SARIMA), measurement of the current MAPE, identification of the difficult item families. 2 weeks.
03
ML model
Training, temporal cross-validation, selection of the optimal model per item family. 4 to 6 weeks.
04
Inventory optimization
Calculation of dynamic safety stocks, reorder points, scenario simulation. 2 to 3 weeks.
05
Integration
Output to ERP / WMS, automatic recalculation, performance dashboard, adjustments based on user feedback.
Get started
Stockouts to reduce or locked-up inventory to optimize?
Tell us about the catalogue size, the target forecast horizon and the source IS. We reply with an initial feasibility analysis within 48 hours.