Analysis of Bank-client Debt Collection Dialogues Using Machine Learning: which Conversational Patterns Increase Debt Collection?

The world's total number of borrowers is gradually increasing. Along with the credit expansion and the debt burden of the population, the risk of default increases. Therefore, the question arises: how to stimulate debtors to return the money? The collection process has become a topic for numerous studies. However, the impact of the operator’s speech in debt collection dialogues has not been sufficiently explored. This study intends to fill the existing gap and identify dialogue patterns in operator speech that positively or negatively affect the level of debt collection. The analysis is based on textual records of telephone conversations of the collection department of one Russian bank with debtors in 1 - 30 days delinquency and attributes of these dialogues. In the course of the research the logistic regression was built on the text data transformed with the TF-IDF algorithm and two nonlinear XGBoost models: the first model was based on the themes revealed with the NMF and the second one on the features of dialogues. As a result, it was found that some speech constructions and attributes of dialogues affect the probability of debt recovery.