Demand Forecasting Model for 3PL Provider

SpeedLog's inventory planners were flying blind. With thousands of SKUs and hundreds of shipments daily, demand was lumpy and hard to predict. The result: stockouts that disappointed customers or overstock that consumed expensive warehouse space.

Building the Forecasting System

We started by analyzing SpeedLog's historical demand data (3+ years). The patterns were clear: strong seasonality (summer peaks, winter valleys), promotional spikes, and day-of-week effects. Traditional ARIMA models captured these patterns reasonably well, but struggled with structural breaks and external shocks.

We built an ensemble: Prophet models captured trend and seasonality, XGBoost models learned complex interactions between promotional calendars and demand, and manual adjustments from demand planners remained part of the process (they knew things the data didn't).

Probabilistic Forecasts

Rather than point estimates ("demand will be 1000 units"), the system produced probabilistic forecasts: 5th percentile (conservative), median (expected), 95th percentile (optimistic). This let inventory managers make risk-aware decisions. When the 95th percentile was very high, they could stock extra. When demand was uncertain, the forecast range told them that.

Integration & Impact

We integrated the forecasts into SpeedLog's inventory management system via daily API calls. The system automatically generated replenishment orders based on forecasted demand and safety stock levels. Human planners could override, but the baseline was now AI-driven rather than intuition-driven.

Within 6 months: forecast accuracy improved from 68% to 94%. Overstock (excess inventory) dropped 35%, freeing up €2M in working capital. Stockouts dropped 68%, improving customer satisfaction. The system paid for itself several times over.