Scope
- To improve forecast accuracy and gain demand insights from secondary sales data for both Consumer goods and Food services
- Leverage Promotion data to fine-tune demand forecast
- Generate Sell-in forecast based on Sell-in Historical Sales data (3 years of data at weekly level) for 2 countries
- Forecast Accuracy measured at Product /Market /Opco’s and month level
- 2 Forecasts generated for M-1 & M-4
- Data trained considering 2 years of history and validate with 1 year of actuals
- Improvement in Forecast accuracy in comparison with current demand planner’s forecast for Month – 4 and Month – 1
- Insights into products with ABC (by volume and value) and XYZ analysis (Coefficient of Variance – Volatility)
Solution
- The forecast was created at the item, item code for Foods Services, distributor level for Consumer Services
- Forecasting was done at the daily, weekly, and monthly level
- Forecasting was done for both M-1 and M-4
- 6 different algorithmic approaches were used for each product, classic techniques such as ARIMA and Holt-Winters as well as machine learning approaches such as CatBoost and FaceBook Prophet
- Engineered features of two different types:
- Statistical features about the time series
- Business features mostly extracted from price and time
Results
- 87% accuracy: Consumer Service, Month-4 Forecast
- 91% accuracy: Consumer Service, Month-1 Forecast
- 81% accuracy: Food Services, Month-4 Forecast
- 73% accuracy: Food Services, Month-1 Forecast
- Machine Learning techniques improves forecast accuracy over the current demand forecaster inputs
- Promotions had a significant impact on the improvement in forecast accuracy in Consumer Goods