Demand forecasting

Distributors, retailers and energy suppliers often run their inventory and procurement on instinct or simple averages. The result: stock-outs that lose sales and frustrate customers, or overstocks that tie up cash and cause write-offs on perishable goods.

The difficulty lies in the hierarchical structure of demand — by product, by store, by region — and its many drivers: seasonality, promotions, public holidays, weather, supply disruptions, local events. Forecasts must stay consistent across levels and granular enough to guide operational decisions.

ADST puts in place hierarchical time-series demand forecasting: we produce forecasts at every level (SKU, store, region, national) and reconcile them to guarantee consistency, so that the sum of store forecasts matches the overall forecast.

The models combine proven statistical approaches with machine-learning methods — gradient boosting on calendar and exogenous variables, deep-learning models for complex patterns. We explicitly incorporate promotions, holidays, weather and cannibalisation effects to capture the true drivers of demand.

Beyond the point estimate, we produce probabilistic forecasts (quantiles) that feed directly into inventory optimisation: safety levels, reorder points and stock-out vs holding-cost trade-offs, tailored to each product by its margin and criticality.