Hierarchical sales forecasting for The Coca-Cola Company - A time series benchmark and initial tool optimization

Hierarchical forecasting with time series has been approached with top-down and bottom-up methods, which have both resulted in satisfying error rates on the most local level. This paper applies bottom-up forecasting to acquire a global sales prediction for The Coca-Cola Company, while assessing the method’s accuracy for multiple large hierarchies. We take the first step in achieving these goals by experimenting with two five-step hierarchies, namely a geographical structure and a product hierarchy, and their combination levels. The first experiments centre around a forecasting tool and its optimization according to the correlation between the short-term prediction horizon and the number of training years that were required for the model. Concretely, the tool seeks to find the optimal combination of hierarchy levels to aggregate over, in order to achieve the most accurate global sales estimate. Furthermore, our findings indicate that there is a limit to the locality that the Holt-Winters triple exponential smoothing model can cope with for global prediction. Subsequently, we delimited the capabilities of this time series analysis for our case study and adjusted the tool accordingly, in order to achieve both accurate and time efficient results. To conclude, a high-level forecast with time series entails delicate balances of the optimization of the model's parameters and of capturing the correct amount of specificity without the unfavourable interference of external factors.status: publishe