Kesintili talep yapısına sahip stok tutma birimleri için birleştirilmiş tahminler

Effective inventory management is vital for the success of companies as it provides a competitive advantage. Many institutions hold different types of stock-keeping units (SKU) on hand. To give an example, SKUs can have a normal pattern like finished goods; have intermittent demand pattern which has no demand for numerous periods and infrequent demand arrivals with highly variable sizes like spare parts. Accurate forecasting is crucial for inventory management. However, forecasting is a challenging task for the SKUs with intermittent demand. The irregular pattern causes traditional forecasting methods like moving averages or single exponential smoothing (SES) to perform poorly. To overcome this problem, several forecasting methods like the seminal method, "Croston's method (CR)," modifications on CR – such as Syntetos-Boylan Approximation (SBA), and new methods like Artificial Neural Networks and Bootstrapping are proposed. Judgmental forecasting methods, which are based on expert opinions, are frequently used in practice as well. These methods are generally used when the historical data is absent and when there are significant changes in the environment and in the time series. Also, it is possible to combine statistical and judgmental forecasts. The studies state that combining forecasts help to eliminate biases and lead to increased accuracy. This study proposes a framework for combining statistical and judgmental methods to generate forecasts with higher accuracy, which at the same time is easy to implement and interpret by the practitioners. First, The SKUs are categorized according to their demand patterns (intermittent, lumpy, erratic, and smooth) by using Syntetos-Boylan-Croston (SBC) categorization scheme. Then SES, CR, and SBA, as one of the variants of CR, are evaluated to select the best method and parameters for each SKU. In parallel, judgmental forecasts should be obtained by using structured methods that reflect the experience and knowledge about the SKUs. Lastly, the generated statistical and judgmental forecasts are combined based on the weighted averaging procedure. The proposed framework is demonstrated on two real-world datasets for spare parts from the automotive and small home appliances industries. The results showed that SES can perform well for SKUs with intermittent demand. Combined forecasts generate higher accuracy than the actual judgmental forecasts of the companies. For most SKUs in the first dataset (automotive), statistical forecasts are the best performing methods. For the second dataset (small home appliances), combination forecasts outperformed both judgmental and statistical forecasts. The outcomes of this study can help institutions to manage their inventory successfully by generating more accurate forecasts. The costs that occur due to stock-outs, overstocks, and risk of obsolescence will decrease, while the customer satisfaction level increases. Also, some spare parts which are strategic for the companies' operations are costly and have a high stock value. Improving the management of such parts will end up with significant savings.