應用層級式預測理論於季節性資料預測 以冷氣機出庫量預測為例

Applying Hierarchical Forecasting Methodology to Seasonal Data Forecast

鍾鎮宇(Chen-Yu Chung)

指導教授 ： 蔣明晃

共同指導教授 ： 郭瑞祥(Ruey-Shan Guo)

國立臺灣大學/管理學院/商學研究所/碩士(2008年)

https://doi.org/10.6342/NTU.2008.10082

摘要

需求資訊的準確性可透過降低需求變異來達成，而層級式預測法中之聚合與分解均為降低需求變異之有效方法。隨產品多樣化及生命週期日益縮短之現象產生，若對一公司中每個產品細項皆進行預測，雖然可以提供企業較為細微之資訊以作決策，但現實面上來看，其所需之實質成本以及時間成本相對龐大，且因為產品生命週期之縮短，亦使得產品細項之預測資訊所能涵蓋有效使用之時間範圍也相對的縮短，以至於無法根據歷史資訊判斷出該產品未來動向及形態。
本研究發現經由導入多變數來修正單變數轉換函數模型後，提升了模型本身解釋及預測能力；再依據分析修正後十五種分解法則所產生之各層級預測值資料發現，表現較好之分解法則為總平均分解法。且於分析各層級最適預測法發現，Top-level最佳之預測法為TT，亦即透過所建構Top-level模型本身預測能力可產生最佳之誤差率；Middle-level最佳之預測法為BM，即透過建構Bottom-level之模型在運用聚合法可使Middle-level產生最佳之誤差率；Bottom-level最佳之預測法為MB，即透過建構Middle-level之模型，再運用總平均分解法使Bottom-level產生最佳之誤差率。且根據分解法分解到向下層級亦發現到，當分解有跨越兩階層時，有放大誤差率之現象。接著經由資料間相關系數之觀察後發現，當資料間具有負相關性，該資料之階層採用聚合法向上聚合可得到較佳之預測值，即此種相關系數判別在季節性資料亦可適用。綜合而言，企業今後如需建立需求預測之模型，可先分析欲預測模型之資料型態，透過欲預測模型資料型態之相關性可判定適合聚合之層級，再透過所分之層級架構之層級數來調整分解法跨越多層級預測有放大現象之缺點，即可決定出所需預測之模型階層為何。如因成本考量限制了預測模型之數量，亦可透過選定總比例分解法，使誤差率不至於偏差過多。此種預先資料的判定，縮減了需求預測之複雜性，即提升了便利性，更可使預測能力更為準確。

關鍵字

層級式預測方法 ； 季節性ARIMA ； 轉換函數 ； LTF

並列摘要

Accuracy of demand information could be lower by decreasing the variance of demand. Hierarchical forecasting methodology is a useful way to decrease the variance of demand. According to the diversification of product and reduction of product life cycle, forecasting all the product items for gaining more information from the items will spend huge cost and time in deriving forecasting model..
In this study we discover multiple variables of transfer function can make the model much accuracy in prediction. According to the revised deposition method, we discover the total average deposition is the best way to deposit data. After analyzing each level, TT (forecasting top-level directly) is the best prediction way in top-level. BM (the prediction value of Bottom-level up to middle level) is the best prediction way in middle-level. MB (the prediction value of middle-level down to bottom-level) by using total average deposition is the best prediction way in bottom-level. When we use TB (the prediction value of top-level down to bottom-level) method cross two levels to generate prediction value, we discover it will increase the prediction error. Observing the correlation of the data, we discover when the data has negative relation we can use the upper integration method to generate better prediction value. Enterprise can observe correlation of the data and deices which level is suited for up-integration method. As considering the cost factor, the enterprise must constrain the forecasting model. The company also can use the best deposition method to decrease the prediction errors. The pre-analysis steps can reduce the complex of deriving the forecasting model and make it much convenience and accuracy.

並列關鍵字

Hierarchical Forecasting ； time series ； ARIMA ； transfer fuction ； LTF

參考文獻

9. 辜莉婷，「應用層級式預測理論於電視購物銷售量預測」，國立台灣大學商

25. Fliedner, G., “An Investigation of Aggregate Variable Time Series Forecast

1. 林威廷，「應用層級式預測方法於季節性資料分析之研究-以家電製造商之出

13. Aigner, D. J. and Goldfeld, S. M., “Estimation and Prediction from Aggregate Data

when Aggregates are Measured More Accurately than their Components”,