- 學位論文
再抽樣法於間斷性需求之預測方法研究
A Study on Forecasting Intermittent Demand with the Application of Resampling Method
摘要
從古自今，無論在財務、行銷、運籌管理、後勤配送等一般日常性作業上，「預測」於企業活動中皆佔有非常重要之地位，一個好的預測方法是促使企業營運更加流暢之關鍵因素。然而，不同型態之時間序列資料擁有不同之特徵，如具快速流通特性之產品 (fast-moving items) 或具緩慢流通特性之產品 (slow-moving items)，因此慎選預測方法將有助於提升預測時之精確性。本研究主要針對具間斷性特性之時間序列發展預測方法，由於間斷性需求與生俱有之特性 (可能包含大量零需求其別)，其將會增加管理者於預測上之困難。本研究嘗試以bootstrap method重複抽樣(resmapling) 之概念並結合馬可夫鏈 (Markov chain) 發展預測演算法，此概念源至於Willemain等人所提出之預測方法，與彼等之主要差異，在於本研究除了考量時間序列中零需求與非零需求之相關性外，亦考量兩個非零需求期別間之相關性。透過自行產生之資料集合，本論文針對我們所提兩種預測方法及指數平滑法、Croston’s method、Modified Croston’s method、Willemain bootstrapping method等共計六種預測方法進行彼此間之精確度比較。研究結果顯示，相較於Willemain bootstrapping method，本研究所提出之預測方法對預測精確度有顯著改善，Croston’s method與Modified Croston’s method亦提供良好之預測品質，然而沒有充分證據可顯示上述六種方法中，任何一種預測方法於不同資料型態均有最好表現。
並列摘要
Forecasting has long been an important technique for enterprises. Most business sectors need forecasting, such as finance, marketing, and logistics. However, different types of time series data exist such as distinct patterns between demands of fast moving items and slow moving items. Hence, selecting an appropriate forecasting model that takes the demand characteristic into account can not be overlooked if we want to obtain a more accurate forecast. This study focuses on a specific type of time series data, called intermittent demand, for which traditional forecasting methods may not be appropriate. In contrast to the forecasting consumer goods, for example, we are likely to be more difficult to accurately predict the demand of spare part due to its nature barrier of intermittent needs. In the thesis, we propose two forecasting algorithms which is essentially based on the concept of bootstrap re-sampling and Markov chain. The idea underlying our method is originated from a patented algorithm proposed by Willemain et al. Differing from the method of Willemain et al , we consider Markov transition probability between periods with zero and non-zero values simultaneously such that we might enhance the precision for forecasting the demand of spare part. Using generated data sets, we compared our algorithms with other methods such as simple exponential smoothing, Croston, modified Croston’s, and bootstrapping. Experimental results demonstrate that our methods have relatively well predictability higher than that of Willemain bootstrapping method. On the other hand, computational studies also revealed that Croston’s and Modified Croston’s methods are quite well in forecasting data with intermittency, but there was no strong evidence to show which method is always the winner.