Judgmental reconciliation of hierarchical forecasts
Producing forecasts at different hierarchical levels (company-level, sector-level, SKU-level) using different aggregations of the data can lead to substantial differences. Various statistical reconciliation approaches have been considered. The bottom-up approach suggests that forecasts should be produced at the lowest level of aggregation (SKU-level) and forecasts at higher levels are derived by simply aggregating the lower-level forecasts. On the other hand, the top-down approach refers to forecasting only the top level and then splitting the forecasts to the lower levels using historical or forecasted proportions. Finally, the optimal approach combines forecasts produced at all levels.
A disadvantage of all these approaches is their full reliance on statistical weighting schemes that do not take into account the special circumstances of each case, thus lacking the judgmental component. Consider the case where different demand planners are responsible for the forecasts produced at the various levels of the company. It is possible that different qualitative information (“soft data”), such as promotional actions or rumours of a new competitive product, is interpreted in a different way from the planners at the different levels. Even worse, some information may not be available at some levels. In such cases, statistical reconciliation approaches would effectively dampen the judgmental interventions made at the various levels.
A way to tackle this limitation is to audit and expand the means of communication and co-operation between demand planners of the various hierarchical levels. In a recent work (forthcoming paper for the International Journal of Forecasting) we suggest the combined use of forecasting and foresight support systems (F2SS). Such systems should bring together common features of forecasting support systems with collaboration and interaction capabilities. A prototype web-based F2SS was introduced in a group of students as an elective exercise in a business forecasting course, showing good levels of satisfaction and influence from team co-operation (that also significantly increased over time).
The use of F2SS is particularly relevant to the problem of judgmental reconciling hierarchical forecasts. Different stakeholders would be able only to share information, but also their views and opinions with regard to the impact of future special events. The use of such systems would render the demand planners able to judgmentally reconcile the differences in the forecasts of the various levels. The result of this approach would be consensus, and not only in terms of numbers! By allowing the forecasters to manually fix any differences in the forecasts, they keep the sense of “ownership” but in a collective manner.
We plan to explore (through a forthcoming judgmental forecasting experiment) the efficiency, in terms of forecasting performance, of a judgmental reconciliation approach for cross-sectional data under the presence of promotions or other special events. Stay tuned!