何時買？買多少？整合潛藏消耗率之層級貝氏預測模型

When and How Many to Buy: A Hierarchical Bayesian Forecasting Model Incorporating Latent Consumption Rates

陳靜怡(Ching-I Chen)

《行銷科學學報》 1卷1期 (2005/04) Pp. 5-21

摘要

購買量與購買時程的預測具有非常重要的策略意涵。購買量的預測是零售商決定舖貨量以及製造商決定產能規劃的依據，購買時程的預測有助於廠商決定行銷策略的執行時點；二者亦是行銷人員評估行銷策略績效的重要指標。過去的研究文獻多半將這兩種購買行為視為獨立變數，而忽略二者之間可能存在的相互依存關係。本文試圖整合存貨消耗率之概念，提出購買量與購買時程之相依預測模型。存貨消耗行為在本質上是無法觀察的潛藏行為，本文將以資料擴充（data augmentation）的觀念建構其模型。除此之外，本文尚透過層級貝氏模型之概念建構三種分析層次之行為模型，得以整合購買時點的行銷活動與日期特性，以及個別顧客特質等各種外在資訊。在實證分析中，本文以國內某油品領導品牌各地加油站之購買紀錄為分析對象，比較層級貝氏模型（HB）與傳統最小平方估計法（OLS）的效度，結果顯示購買時點層次HB估計元之參數估計能力與樣本內預測能力最佳。

關鍵字

存貨消耗模型 ； 層級貝氏模型 ； 資料擴充

並列摘要

Prediction of purchase quantity and timing has very important strategic implications. Purchase quantity forecasting provides a criterion for retailers' assortment strategy and manufacturers' production planning; purchase timing forecasting facilitates firms to decide when to put strategies in practice. The effects of marketing activities on purchase quantity and timing are also important criteria to evaluate marketing performance. The two purchase behaviors were often viewed as independent but not interdependent response variables in previous related literature. This paper attempts to construct a forecasting model of the two interdependent purchase behaviors incorporating the concept of the inventory consumption model. However, the inventory consumption behavior of customers is unobservable in nature, so this paper employed the concept of data augmentation to model this unobservable response variable. Besides, this paper adopted hierarchical Bayesian approach to combine three analysis levels of models to incorporate each kind of information from data, including the inventory consumption model on unit timing level, the purchase quantity and consumption rate model on purchase timing level, and the marketing effects model on individual customer level. Empirically, we employed the petrol purchase record database covering all gas stations of a domestic leading petroleum brand to compare the validity of HB approach to traditional ordinary least square estimators (OLS), and the empirical result showed that the HB estimators of purchase timing level had the best predictive ability.

並列關鍵字

Inventory Consumption Model ； Hierarchical Bayesian Model ； Data Augmentation