A column generation approach for product targeting optimisation within the banking industry

Abstract

Product targeting optimisation within the financial sector is becoming increasingly complex as optimisation models are exposed to an abundance of data-driven analytics and insights generated from a host of customer interactions, statistical and machine learning models, and new operational, business, and channel requirements. However, given the expeditious change in the data environment, it is evident that the product targeting formulation cited throughout the literature still needs to be updated to align with the realistic modeling dynamics required by financial institutions. In this thesis, an enhanced product targeting formulation is proposed that incorporates a large set of new modeling constraints and input parameters to try and maximise the economic profit generated by a financial institution. Furthermore, the proposed formulation ensures that the correct product is offered to the desired customers at the best time through their preferred communication medium. To solve the preceding product targeting formulation, a novel column generation approach can reduce problem complexity and, in turn, allow for significantly larger problems to be solved to global optimality within a reasonable time frame. The column generation approach proposed in this thesis allowed the solution of complex product targeting formulations up to problem sizes consisting of 25000 customers, 35 products, and three channels. On the other hand, the standard branch-and-bound algorithm of Cplex could only solve problem instances up to a size of 5000 customers, 20 products, and 3 channels. The results show that the proposed column generation framework can significantly reduce the memory requirements of the various test instances, allowing it to solve significantly larger product targeting problems compared to standard solution methodologies. Furthermore, the solution framework reached close to global optimality for most of the test cases considered in this thesis.