Forecast reconciliation in the temporal hierarchy: Special case of intermittent demand with obsolescence

Abstract

Intermittent demand items are often exposed to a risk of obsolescence, which translates into a downward trend if related series are observed at reduced frequencies. However, the existing forecasting applications in the said context are limited to the modifications in the Croston method. Moreover, they ignore the potential benefits of utilising the information present at reduced frequencies. Accordingly, the study proposes using the temporal hierarchy (TH) framework to model the decreasing trend information for intermittent demand items efficiently. Additionally, with various optimal reconciliation techniques under the TH framework, the study attempts to provide an in-depth understanding of their expected performance in the obsolescence scenario.Further, the study argues that the reconciled forecast at the bottom level for these techniques can be expressed as a combination of disaggregated forecasts. Wherein the scheme of the combination depends on the set of assumptions followed by each approach. However, their application is limited by the tendency to produce negative reconciled forecasts, which is of particular concern for intermittent demand items. Therefore, the present study proposes using the constrained least square estimation framework to determine the desired optimal combination scheme for the bottom-level forecasts, which are then utilised to enforce coherency across the considered hierarchy. Besides, the study corroborates the suitability of the proposed framework in the obsolescence scenario with the help of simulated and empirical intermittent demand datasets. Finally, the results are validated with the help of relevant statistical analysis.