Data-driven First Touchpoint Scheduling in InsureTech Domain
- Lead calling is one of the foundational promotional activities that bring business the first human-to-human interaction with potential customers. A successful call can build the root of trust between the business and potential customers, which ultimately converts the call into sales. However, conducting lead calls faces big challenges as customers nowadays are being bombarded by sales calls due to the unsustainable calling strategy of many businesses. To conduct lead calls in a sustainable and effective way, this thesis builds a first touchpoint scheduling system based on the lead call history of the InsurTech company wefox Group. By applying a set of machine learning techniques, the system can generate a list of leads with the highest reaching probabilities for the period requested by call agents. Thus, leads are called at the timing that is convenient to answer, meanwhile, call agents can work more productively with a higher reach rate.
Exploratory data analysis is conducted to understand the relationship between the reach status and the customer data. 14 classifiers are applied to predict the reaching probability. Brier Skill Score, Average Precision Score, Precision-Recall Curve, Precision, Recall, fit time and a customized evaluation metric are selected to evaluate the classifiers’ performance. Four resampling methods, class weight adjustment and a meta-model to adjust the class threshold are used to deal with class imbalance. Bayesian Optimization is used for hyperparameter tuning. The combination of mutual information classifier and correlation feature selection approaches helps reduce the number of features needed.
The experiments show that LightGBM classifier outperforms the other classifiers with Brier skill score of 0.181, average precision of 0.487 and recall of 0.642. Feature importance indicates that besides the customer interaction record with business, macro factors such as unemployment rate, social assistance rate are also powerful features. Business value evaluation indicates that the system can improve the time effectiveness of call agents by 158%. 72% of records have no overlapping between actual non-reachable hours and predicted reachable hours.
MetadatenAuthor: | Yuxin Wang |
---|
URN: | urn:nbn:de:kobv:b721-opus4-25917 |
---|
Advisor: | Roland Müller, Anish Arora |
---|
Document Type: | Master's Thesis |
---|
Language: | English |
---|
Date of first Publication: | 2021/02/23 |
---|
Publishing Institution: | Hochschulbibliothek HWR Berlin |
---|
Granting Institution: | Hochschule für Wirtschaft und Recht Berlin |
---|
Release Date: | 2022/01/11 |
---|
Tag: | Brier Skill Score; LightGBM; class imbalance; lead calls; probabilistic classification |
---|
Page Number: | 94 |
---|
Institutes: | FB I - Wirtschaftswissenschaften |
---|
Licence (German): | Urheberrechtsschutz |
---|