利用BG/NBD模型預測顧客未來行為，以提升流失管理效率--以藥品產業為例

Abstract

本研究利用顧客交易資料，包含顧客重複購買次數、最近一次購買與購買歷史總時間，透過BG/NBD模型(Fader et al., 2005)來預測顧客未來交易次數與顧客活躍機率。此研究可利用模型預測結果，對顧客作一對一行銷方案，將高交易金額的顧客與其活躍機率作比對，找出需要重新制定新行銷方案的顧客，以增加顧客權益與提高顧客流失管理效率。

本研究之資料為台灣某人類肝臟用藥的交易資料，交易時間為97年1月至98年2月共六十週的交易資料，本研究將其資料分為兩部份，前30週作為建模資料，後30週資料作為檢驗模型準確度。交易顧客為非契約性的醫院、診所與藥局，總共為953個顧客。非契約顧客之研究主因為其非購買行為難以判斷為交易關係流失或者只是在兩次交易時間的暫時停止交易，因此本模型計算出的顧客活躍機率能計算每位顧客於特定時間其交易存活機率。

本研究結果，根據顧客重複購買次數的預測值與實際值，其MAPE (mean absolute percentage error) = 42.6% < 50%為可靠預測；且透過變異數分析，顧客活躍機率之高低對顧客總金額有顯著影響，而顧客活躍機率之高低對顧客購買的平均金額並無顯著影響。

In the study, we use a transaction database containing information on the frequency and timing of transactions for a list of customers in order to forecast about future purchasing. The forecasting results help us develop one to one marketing plans to increase customer equity and churn management efficiency. The data for a sample of 935 non-contractual customers who made their purchase of some particular hepatic capsule with the pharmaceutical company is during the first 30 weeks of 2008. We have information on their repeat purchasing behavior up to the lasting 30 weeks. The first 30 weeks purchasing data is used for modeling. The last 30 weeks information is used to measure accuracy. The types of customers are hospitals, clinics and pharmacies in Taiwan. The result shows the model reliable for the mean absolute percentage error 42.6% < 50%. The level of customer active probability affects customer transaction total monetary significantly by analysis of variance, while the customer active probability has no significant effect on customer average transaction monetary.