層級資料結構下預測演算法之效能比較

Abstract

在現今全球化之供應鏈運營中，具有來自跨通路部門及市場需求，為確保供應鏈可平穩高效地運行，需分析來自不同部門的預測，並使用時間序列模型在連續時間獲得的數量的有序量值，如產品銷售量、股票價格、每日溫度測量和每週票房數據等資料，而當前時間序列模型以ARIMA、ETS等模型為主流，以預測數列間的季節性、趨勢性以及隨機性。 分層時間序列 (HTS) 是遵循分層聚合結構的時間序列的集合。假設有一家食品供應商共有數十種產品與口味。其中各個產品的銷售即為一時間序列。各個通路銷售的食品個數亦為一個時間序列，而這些時間序列的集合具有層次聚合結構，彼此數量間皆有一定程度的相關，分層時間序列用以探討其中的相關性並提供更佳的預測方法。 本篇論文主要探討關於各個SKU的需求性質，並從中疏理分層時間序列演算法較可準確預估各產品需求量之情境，目前較常見之分析方法Bottom-up approach、Top-down approach、Reconciliation approach 等演算法，本研究將自各種SKU需求性質判斷需要採用之演算法，以預測值與現實值之RMSE、MASE作為主要演算法損失函數，並從中減少預測誤差以降低成本。

In modern global supply chain operations, there are demands from cross-channel departments and markets. To ensure that the supply chain can run smoothly and efficiently, it is necessary to analyze forecasts from different departments. With time series models we could obtain continuous-time quantitative estimates of ordinal values such as product sales, stock prices, daily temperature measurements, and weekly box office data. Hierarchical time series are collections of time series that follow a hierarchical aggregation structure. For example, imagine a food supplier with dozens of products and SKUs. The sales of each product are time series. The numbers of foods sold by each channel are time series as well. And the collection of these time series has a hierarchical aggregation structure. This study will discuss the demand properties of each SKU, and analyze the scenario in which hierarchical time series algorithms, such as the Bottom-up approach, Top-down approach, and Reconciliation approach, could accurately estimate the demand of each product in different nodes. This study shows how to determine the algorithm that needs to be adopted based on the properties of various flavor demands, the RMSE of the predicted value and the actual value is considered as the main algorithm loss function, and the prediction error is reduced to further lower the cost.