An aggregate–disaggregate intermittent demand approach (ADIDA) to forecasting: an empirical proposition and analysis

Abstract

Intermittent demand patterns are characterised by infrequent demand arrivals coupled with variable demand sizes. Such patterns prevail in many industrial applications, including IT, automotive, aerospace and military. An intuitively appealing strategy to deal with such patterns from a forecasting perspective is to aggregate demand in lower-frequency ‘time buckets’ thereby reducing the presence of zero observations. However, such aggregation may result in losing useful information, as the frequency of observations is reduced. In this paper, we explore the effects of aggregation by investigating 5000 stock keeping units from the Royal Air Force (UK). We are also concerned with the empirical determination of an optimum aggregation level as well as the effects of aggregating demand in time buckets that equal the lead-time length (plus review period). This part of the analysis is of direct relevance to a (periodic) inventory management setting where such cumulative lead-time demand estimates are required. Our study allows insights to be gained into the value of aggregation in an intermittent demand context. The paper concludes with an agenda for further research.

Keywords:

• demand forecasting
• inventory management
• intermittent demand
• aggregation
• empirical investigation

Acknowledgements

The work conducted by Aris Syntetos and John Boylan has been funded by the Engineering and Physical Sciences Research Council (EPSRC, UK) grant no. EP/F012632/1 (a project entitled: Forecasting and inventory management: bridging the gap). More information on this project may be obtained at: http://www.mams.salford.ac.uk/CORAS/Projects/Bridging_the_Gap/.

Notes

1 For presentation purposes, and for the remainder of the paper, the ADIDA process will be denoted by: ADIDA (aggregation level, extrapolation method, disaggregation method).

2 95% confidence intervals were constructed at aggregation level=1 through the calculation of the sample standard errors for both methods. Subsequently, all average errors at the various aggregation levels (>1) indicated in Figure 4 were evaluated as to whether they constitute statistically significant improvements/reductions.