A number of research projects have demonstrated that the efficiency of inventory systems does not relate directly to demand forecasting performance, as measured by standard forecasting accuracy measures. When a forecasting method is used as an input to an inventory system, it should therefore always be evaluated with respect to its consequences for stock control through accuracy implications metrics, in addition to its performance on the standard accuracy measures. In this paper we address the issue of judgementally adjusting statistical forecasts for ‘fast’ demand items, and the implications of such interventions in terms of both forecast accuracy and stock control, with the latter being measured through inventory volumes and service levels achieved. We do so using an empirical dataset from the pharmaceutical industry. Our study allows insights to be gained into the combined forecasting and inventory performance of judgemental estimates. It also aims to advance the practice of forecasting competitions by arguing for the consideration of additional (stock control) metrics when such exercises take place in an inventory context.
In many organisations, the size and complexity of the demand forecasting task at the individual Stock Keeping Unit (SKU) level necessitates the use of statistical methods, such as exponential smoothing, instead of judgemental methods. However, in many cases these statistical forecasts will be subject to judgemental adjustments by managers (Fildes and Goodwin, 2007 and Sanders and Manrodt, 1994). This is true for both ‘fast’ and ‘slow’ demand items. Some academic research has been carried out to ascertain the effectiveness of these adjustments with respect to forecast accuracy (e.g., Fildes, Goodwin, Lawrence, & Nikolopoulos, 2009 and Mathews & Diamantopoulos, 1992). However, the stock control implications of these human judgements have not received sufficient attention. We have recently undertaken a project (Syntetos, Nikolopoulos, Boylan, Fildes, & Goodwin, 2009) where we explore the inventory performance of adjustments when they are applied to statistical forecasts of intermittent demand. In this paper, we extend this analysis to the case of fast demand items by means of experimentation with an empirical dataset. We argue that forecasting performance in an inventory context should always be evaluated with respect to its implications for stock control, through accuracy implications metrics, in addition to considering forecast accuracy measures.
Similar arguments have been advanced in the academic literature in the contexts of different forecasting applications. In the area of economics and finance, for example, Timmermann and Granger (2004) highlighted the need to evaluate forecast results using utility functions. Often the predictive approach that is best based on a given accuracy metric will not be the one that outperforms competitors if utility measures are employed, such as financial outcomes, inventories, customer satisfaction, or socio-economic benefits.
The remainder of our paper is structured as follows: in the next section we discuss the importance of assessing the stock control performance of estimators in an inventory context. In Section 3 we provide details of the empirical dataset available for the purposes of our investigation and of the structure of our simulation experiment. Section 4 presents the empirical results and their interpretation, followed in Section 5 by the conclusions drawn from this work, along with some suggestions for further research in this area.
The distinction between forecast accuracy measures and accuracy implication metrics has already been established in the literature. However, the scale of inventory volume reductions that may result from relatively modest improvements in forecasting accuracy has not been examined in detail. In this study we have shown that reductions in MAPE and sMAPE of the order of 1% may be translated into inventory reductions of the order of 15%–20%. At the same time, cycle service levels and fill-rates are improved by approximately 1%.
These results are significant for forecasting theory and practice alike. Firstly, it has been shown that judgemental adjustments have the potential to improve forecast accuracy over longer horizons than those previously demonstrated by Fildes et al. (2009). Secondly, as noted above, these improvements translate to substantial inventory reductions. This shows that investments in intelligent judgemental adjustments may reap considerable financial returns.
This research is based on a relatively small dataset. We have attempted to analyse the results in relation to demand characteristics, but no valuable information was produced. However, a larger sample of SKUs would facilitate such an analysis. For example, the performance of the judgemental adjustments could then be evaluated for different classes of SKUs in an ABC-type classification by volume. The availability of cost information would also have enabled a detailed assessment of the financial implications of adjusting statistical forecasts, in addition to further exploring linkages to SKU characteristics, say through ABC classifications by value. Moreover, it would also be beneficial to replicate this study on other companies and industries. Finally, more theoretical work is needed to understand the subtle linkages between forecast accuracy measures, such as MAPE and sMAPE, and accuracy implication metrics, such as inventory holdings and service levels.
