Our client is a global consumer packaged goods company working in the food product space. The forecasting process they currently have in place is extremely effort-intensive and driven by heuristics. The client wanted to have a robust and transparent demand planning process that is consistent across all markets and categories.
The objective was to
- Develop forecasting models to predict shipments with reasonable accuracy for three years
- Provide a breakdown of the predicted shipments into baseline and incremental components
- Develop a scalable solution framework to forecast shipments for multiple products across all markets
- Huge volume of data with millions of records and thousands of features
- No clear view of base demand and lift due to various factors
- Custom treatment of different types of products (like discontinued, seasonal, new or converted items)
- Handling the product x retailer combinations with sparse data (no frequent shipments)
- Derived features from internal and external data. Some of the features are – seasonality & trends, the effect of holiday & sports events, price & promotion, retailer purchasing patterns, product purchasing patterns, POS consumption patterns, macro-economic indicators, weather, etc.
- Built data pipeline in Hue – Impala, CDSW for modeling
- Used an ensemble of Random Forest and FB prophet for model development based on benchmark performance and explainability of the forecast
- Measured Model performance: calculated MAPE & Bias and compared at various granularities
- The shipment forecasting solution developed brought about an estimated revenue uplift of USD 80 M – USD 100 M
- Improved forecast accuracy by 5%-20% and reduced the time to implement by 60%