The Challenge

A regional wholesale distributor managing 15,000+ SKUs across 8 warehouses was struggling with demand forecasting. Their Excel-based forecasting models relied heavily on historical averages and manual adjustments, leading to frequent stockouts of high-demand items and excess inventory of slow-moving products.

Forecast accuracy was around 68%, resulting in lost sales, emergency shipping costs, and $4.2M in excess inventory write-downs annually. The sales and operations teams spent 40+ hours weekly manually adjusting forecasts based on gut feel.

Our Solution

We built a machine learning forecasting engine that combines multiple prediction models and external data sources:

  • Time series models (ARIMA, Prophet) for stable products with consistent demand patterns
  • Gradient boosting models (XGBoost) that incorporate dozens of features: seasonality, promotions, weather, local events, economic indicators
  • Deep learning models (LSTM) for products with complex, non-linear demand patterns
  • Ensemble methods that automatically select the best model for each product-location combination
  • Real-time demand sensing that adjusts forecasts based on incoming orders and external signals

The system automatically generates forecasts at the SKU-warehouse-week level, flags anomalies for human review, and provides confidence intervals so planners can make risk-adjusted decisions. It integrates directly with their ERP system and updates forecasts nightly.

Results

32% improvement in forecast accuracy: Overall forecast accuracy increased from 68% to 90%, with the biggest gains in seasonal and promotional products.

$2.4M annual inventory cost savings: Reduced excess inventory by 35% and stockout-related lost sales by 60%, driving significant bottom-line impact.

85% reduction in manual forecasting time: Planning team now focuses on strategic decisions and exception handling rather than spreadsheet maintenance.

Faster response to demand shifts: The system detected and adapted to COVID-19 demand changes 2-3 weeks faster than the old manual process.

Technology Stack

Python with scikit-learn, XGBoost, TensorFlow, and Prophet for modeling. Apache Airflow for orchestration, PostgreSQL for data storage, integrated with SAP ERP via REST APIs. Dashboard built with React and Plotly for interactive forecast exploration.

Client Testimonial

"The forecasting engine has been a game-changer for our operations. We went from dreading monthly forecast meetings to actually trusting our numbers. The system caught demand shifts during COVID that we would have missed for weeks with our old approach."

— Jennifer Park, VP of Supply Chain

Discuss Your Project