انتخاب روش پیش بینی در یک زنجیره تأمین جهانی

In supply chains, forecasting is an important determinant of operational performance, although there have been few studies that have selected forecasting methods on that basis. This paper is a case study of forecasting method selection for a global manufacturer of lubricants and fuel additives, products usually classified as specialty chemicals. We model the supply chain using actual demand data and both optimization and simulation techniques. The optimization, a mixed integer program, depends on demand forecasts to develop production, inventory, and transportation plans that will minimize the total supply chain cost. Tradeoff curves between total costs and customer service are used to compare exponential smoothing methods. The damped trend method produces the best tradeoffs.

A comprehensive review of research in forecasting for supply chains is given by Fildes and Kingsman (2010), who conclude that there are few findings of any managerial importance. We agree. To ensure mathematical tractability, most researchers have assumed greatly simplified operating systems and cost structures. Furthermore, most have also failed to match the generation process for demand with the choice of a forecasting method. Thus, forecast errors have been compounded with misspecification errors, making it difficult to understand the effects of forecasting on efficiency, costs, inventory investment, or customer service levels. In a careful MRP simulation, Fildes and Kingsman set out to correct many of the fallacies in previous research. They found that the benefits of improved forecasting are considerably greater than the effects of choosing inventory decision rules, and that a misspecification of the forecasting method leads to increases in costs. Fildes and Kingsman call for more empirical modelling of the supply chain that is grounded in observed practice, and that is the theme of this paper. We model the relationship between forecasting and operational performance in the supply chain of a global manufacturer of lubricants and fuel additives, products which are usually classified as specialty chemicals. The model includes four manufacturing plants and daily time series of actual demand collected over a four-year period. Both optimization and simulation techniques are used to develop production schedules, inventory targets, and transportation plans for shipments between plants and to customers. Optimization depends on demand forecasts, supplied by exponential smoothing, and is done with a mixed integer program in order to minimize total variable supply chain costs. Management asked for forecasting methods that were simple and easily automated, making some form of exponential smoothing the only reasonable choice. We considered three methods: simple exponential smoothing (SES), Holt’s additive trend (Holt, 2004), and the damped additive trend (Gardner & McKenzie, 1985). SES and the damped trend are obvious choices, given their long record of success in empirical studies (Gardner, 2006); the data suggested that the Holt method would not perform well, but it was retained as a benchmark for the other methods. To select the best method, tradeoff curves were computed between total supply chain cost and several measures of customer service. The damped trend gave the best operational performance for any level of cost, followed by SES and Holt. It is interesting to contrast these results with traditional method selection based on average forecast accuracy measures; surprisingly, SES gave the best overall average accuracy.

There appears to be no previous research on forecasting method selection based on operational performance in a real supply chain. The supply chain model in this paper is driven by actual daily demand data and integrates exponential smoothing, optimization, and simulation. We show that the choice of forecasting method makes a significant difference to both the customer service and cost tradeoffs available to management. Hyndman and Koehler’s scaled error measures are the best available options for measuring the average forecast accuracy, but there is no relationship between operational performance and average accuracy across all products in this supply chain. Syntetos, Nikoloupoulos, and Boylan (2010) argue that, in comparisons of average accuracy for inventory demands, the errors should be weighted by the cost or customer service impact. We agree that this should be done in a pure distribution inventory, where there are few interactions between inventory items, but it is difficult to do so when modeling production, transportation, and distribution in the supply chain context. The consequences of forecast errors are complex because there are powerful interactions between products competing for the same production capacity. These interactions lead us to the conclusion that forecasting must be evaluated at the aggregate level in the form of cost-service tradeoff curves for the entire supply chain. Finally, one obvious question about this research is whether the damped trend is superior to the company’s existing forecasting method. Anecdotal evidence suggests that the damped trend is an improvement, but the company has no clearly defined forecasting method at present, so there is no real basis for comparison. The company has relied on purely subjective forecasts for many years, and there are no reliable records of forecast values, when forecasts were made, or how they were made. Our experience is that this is not unusual.