Datadriven retention i relationsintensiva tjänstekontexter : Möjligheter och begränsningar för churn-prediktion inom Private Banking

Sammanfattning

As firms adopt data-driven decision-making, predictive modeling has gained an increased importance in the work with customer retention. However, its value in relationship-intensive contexts remains unclear, as customer loyalty in such settings appears to depend on relationships, soft values, and aspects that require tacit knowledge to interpret, rather than just factors that are observable in explicit data. This study investigates whether predictive churnmodeling can create value beyond advisors’ professional judgement in Private Banking. The study is conducted as a case study at a Swedish Private Banking firm and applies a pragmatic approach to examine the practical value of predictive modeling beyond established statistical performance metrics. The study combines quantitative model development, through the construction and evaluation of XGBoost models, with qualitative interviews with eight advisors.The models and their statistical performance metrics are mainly used to examine whether churnsignals can be captured in customer and business data within a relationship-intensive context, while interviews with advisors are used to nuance these statistical metric results and assess whether such modeling can provide additional value in relation to advisors’ customer knowledge, work practices, and understanding of customer relationships. The findings show that the models have a certain but limited predictive ability in this context. As churn is a rare event, the data is characterized by substantial class imbalance, which makes it difficult to achieve high precision. However, the models achieve a top-decile lift of approximately four, indicating that the customer groups identified as high risk have a churn rate about four times higher than a random selection. The findings further indicate that predictive churnmodeling can create value in Private Banking, not by replacing advisors’ professional judgement, but by complementing it. The model’s strengths partly correspond to advisors’ limitations, such as creating overview and consistently screening large customer groups, while advisors are better able to interpret relational, contextual, and tacit signals that are not captured in explicit data