The project goal was to optimize BASF’s global supply chain and production line for custom materials. Due to the varied nature of parts used in the automobile industry, different models of cars require unique pieces. As car models are discontinued, demand for their parts decreases. Further, the intermittent nature of custom material demand and spare parts meant that BASF must constantly worry about potential capacity breaches. Considering a lead time from order to ship-out of two months, our aim was to give BASF’s automobile division a monthly demand forecast for the next two months for each unique material.
The original data extracted through the BASF ERP system contained the number of daily shipped units for 826 unique materials from its automotive division. Given our forecasting goal, we aggregated the irregularly-spaced demand units into regular monthly intervals. Materials with negative and zero demand volumes in the test set period have been filtered out. The training period was 59 months long (10/12-8/17); the test period 12 months (09/17-08/18). RMSE in the test period was used as the performance metric as per BASF company standard evaluation.
In order to provide a simple heuristic to understand both the behavior and importance of different materials, the material series were grouped using ABC-XYZ analysis. For all series in each cell of the ABC-XYZ matrix, we tested commonly-used forecasting methods and selected the method producing the lowest average RMSE, while also taking into account computing time. The chosen model was then used to forecast the future demand for each material in that group. In nearly all groups, a naive forecast was found to be best, meaning past order quantities were often the best indicator for future orders. Exponential Smoothing (ETS) performed better than the naive in groups AZ and BZ, which represent the series with lowest forecastability and high to medium importance. In the low forecastability, high-importance group AY, an ensemble of methods represented a nearly 19% improvement in RMSE over the naive. Given the global scale of BASF’s operations, these relatively minor improvements could potentially save millions of dollars in excess inventory costs.
As the majority of factory and production managers are not data analysts, we built an easy-to-use interactive forecasting platform using R Shiny to allow simple but immediate forecasts of quantities per series. The platform makes forecasts (both point and interval) based on our recommended method for each group of materials and shows these forecasts in relation to an estimated maximum capacity. The system further provides a historical view of a material’s demand behavior and an interface for finding materials with similar historical demand characteristics to give an idea of future behavior.
Going forward, we recommend that BASF consider the unit cost of materials when conducting the ABC- XYZ analysis. We also suggest that BASF provide more information related to the types of materials to more easily identify correlated time series. Lastly, we advise future teams to explore the performance of forecasting methods using a multi-year, roll-forward validation scheme.