본 연구에서는 일반 소비재 제조업이나 유통업과 같이 이탈예측이 상대적으로 까다로운 비계약업종(Non- Contractual Business)에서의 효과적인 이탈고객에 대한 정의와 예측을 가능하게 하는 세 가지의 대수적 모형을 제시하였다. 첫 번째 모형은 고객 개인마다의 평균적인 구매주기를 고려할 수 있는 개인평균구매주기(API: Average Purchasing Interval) 모형, 두 번째 모형은 고객 개인별 구매주기의 최대치를 고려할 수 있는 개인최대구매주기(MPI: Maximum Purchasing Interval) 모형, 마지막으로 세 번째 모형은 고객의 자연구매가 발생할 수 있는 최대의 전제기간을 고려할 수 있는 개인허용구매주기(TPI: Tolerance Purchasing Interval) 모형이다. 이와 같이 고객들의 상대적인 구매주기 패턴을 고려한 대수적 모형들의 효용성을 검증하기 위해 전체 고객의 평균구매주기를 활용한 이탈예측 모형, BG/NBD 모형 등 고객의 과거 구매행태 정보를 토대로 고객들의 이탈을 예측하고 있는 기존의 비계약업종에서의 이탈예측 모형들과 성과를 비교하였다. 모형 간의 성과를 비교 검증하기 위해 이탈 예측 정확도(Precision) 및 재현율(Recall)을 이용하였으며, 뿐만 아니라 재무적인 관점에서의 효익 측면에서도 모형 간 예측력의 유의미함을 검증하였다. 분석 결과 세 가지 이탈예측 모형 중 개인허용구매주기(TPI) 기준의 이탈예측 모형이 가장 높은 예측 정확도(66.03%)를 보이고 있는 것으로 나타났다. 예측 정확도는 개인평균구매주기(API) 기준 모형에서 개인최대구매주기(MPI), 개인허용구매주기(TPI)순으로 점차 개선되는 것을 볼 수 있었으며, 실제 이탈고객에 대한 모형의 재현율은 이 순서로 오히려 감소하고 있는 것을 볼 수 있었다. 즉, 각 모형 간에는 예측 정확도와 재현율 사이에 상쇄관계(Trade-Off)가 존재 하는 것을 볼 수 있는데, 이는 개인평균구매주기(API) 보다는 개인최대구매주기(MPI)가, 개인최대구매주기(MPI) 보다는 개인허용구매주기(TPI)가 개인 이탈에 대한 기준을 보다 엄격하게 적용함으로써 이탈이라고 판단하는 예측 고객의 모수 자체가 작아지기 때문이라고 볼 수 있다. 따라서 어떠한 이탈정의 기준의 예측 모형을 적용할 것인지는 이탈예측 모형의 예측 정확도와 재현율 정도에 따라 파생되는 비용 혹은 이익의 증감 정도를 파악하여 적용하는 것이 바람직할 것이다. 또한 이와 같은 분석 결과를 토대로 산업별 특징에 알맞은 이탈정의 기준을 제시하였으며, 더 나아가 상대적 구매주기 패턴을 고려한 이탈예측 모형 활용 방안으로 이탈 가능성 단계별로 차별화된 이탈관리를 수행할 수 있는 이탈방지 전략인 고객이탈 조기경보체제를 제안하였다. 비계약업종의 평균적인 이탈정의에 대한 고객이탈 예측 정확도가 30% 내외인 것을 감안할 때, 고객들의 상대적인 구매주기 패턴을 고려한 이탈정의 기준의 예측 모형은 CRM 전략을 수행함에 있어서 유의미한 도구가 될 수 있을 것으로 기대된다.
It is no doubt that prediction and prevention of customer defection is one of most crucial CRM activities. However, most previous researches have been studied in contractual business areas such as insurance and telecommunications because the operational definition of customer defection is comparatively simple; expiration or cancelation without re-contract means customer defection in those businesses. Meanwhile, since such obvious severance behaviors are scarcely appeared in non-contractual business, prediction and prevention of customer defection in such business types are irrelevant. That`s why only few customer defection studies have been addressed in non-contractual business. Besides, there exists a critical limitation in customer defection models addressed in such studies. They usually define customer defection based on recency only, e.g., customers not purchased in recent six months, which neglects the situation of voluntary returns. Voluntary return is the phenomenon that the customers who were expected to be churned come back voluntarily. If customer defection was determined only with recency regardless of such characteristics, the effectiveness of churn prediction model might be disappointed no matter how the model is elaborate or sophisticated. To fill the deficiency and surmount the limitation of customer defection studies in non-contractual business area, we designed three different algebraic models to decide customer defection by using individual purchasing pattern, and suggested a systematic model for defining and predicting customer defection appropriate to non-contractual businesses through comparing the three algebraic models. The first model was based on individual average purchasing interval. In this model, average purchasing interval(API) that is the average of purchasing intervals of specific customer and last-purchasing lapse(LPL) that is the inverse value of recency should be calculated. After comparing two values, API and LPL, the customer whose LPL value is bigger than his or her API value would be regarded as a churned customer, otherwise not. This model means that a company should not apply a single cut value of recency (e.g., six months) that might be calculated through averaging all the customers` purchasing intervals to judge customer defection, rather they have to withhold to declare some customers to be churned until reaching their own API value by considering individual purchasing patterns. The second model is quite similar with the first one but has different purchasing interval value: maximum purchasing interval(MPI) that is the maximum value of purchasing intervals of specific customer. Similarly with the first one, MPI value is compared with LPL value to decide whether or not the customer is defected; the customer would be judged as a churned customer if his/her LPL value is bigger than his/her MPI value. This model stresses that the waiting period for voluntary return should be extended not to API but to MPI in due consideration of customer`s personal purchasing histories. The last one is the tolerance purchasing interval(TPI)-based model: which is considering till when a company should tolerate as the likelihood duration of spontaneous buying. According to this model, we first derive the maximum purchasing deviation(MPD) which is the biggest deviation of purchasing interval from API. Then, we calculate the TPI value from adding the MPD into the API. The customer would be judged as a churned customer if his/her LPL value is bigger than his/her TPI value. After comparing these models with the recency-based model and the probabilistic ones such as BG/NBD, we finalized that the proposed models have higher performances in terms of precision, recall, and cost-benefit perspective. We analyzed a data set of observation window spanned from June 1, 2008 through May 31, 2010 of the biggest retailer in Korea. We extracted first the data of customers from the huge data set because the prevention of customer defection is the most important marketing strategy to the highly-valued customers in this type of industry. The sales data of June 1, 2008 through May 31, 2009 was used to build the models(training data set), and the other half of data set was used to validate the models(test data set). To determine whether the models predict customer defections correctly, we discerned the defected customers only when the customers who had a purchasing history in the first half of durations(June 1, 2008 to May 31, 2009) but did not have a purchasing history in the second duration(June 1, 2009 to May 31, 2010). The results show the TPI-based model has the highest precision(66.03%) but the lowest power(recall) of prediction. This is a trade-off relationship caused by TPI has more strict criterion to say "you seem to be defected" than API or average purchasing interval of total customers, consequently deriving comparatively few customers as defected from the population. Therefore, we recommend that companies should adopt a model by considering their marketing costs and expected profits subjected to the anticipated churned customers. For example, to manage a VIP customer group that is probably to need more cost and time than a normal customer group does, they had better adopt the MPI or TPI-based model because it would draw only few defected customers but more prudently and correctively. Contrarily, the API-based model would be more proper to a bigger pool of targeted customers, which usually consumes less resource to prevent their defection and has no critical issue of correctness. Considering that average precision of churn prediction in non-contractual businesses was around 30%, the definition model of customer defection based on individual purchasing intervals might increase the effectiveness of churn prevention strategy. To step forward from these results, we developed a practical framework of early-warning system for customer defection, which can signal the level of risk to churn by stages. Companies can utilize this early warning system so that they would regard the strategy of customer defection management as not a deterministic judgment but a stochastic risk management.