PRIME-RE: A Parallel and Recursive Requirements Elicitation Approach for ML-Based Financial Decision Systems

With the growing use of machine learning models for supporting financial decisions, financial institutions face increasing pressure to provide explanations, traceability and stable risk assessments of their decision-making processes. Current approaches to requirements gathering do not take into account the realities of this domain since they are based on the assumption that decision requirements can be formulated at an early stage and then be refined incrementally. In reality, however, financial decision objectives; data availability; analytical feasibility; and process governance constraints evolve simultaneously and interactively and are often discovered only by way of exploratory analyses or preliminary validation activities. This paper introduces PRIME-RE, a parallel and recursive methodology for eliciting requirements for machine-learning based financial decision support systems. PRIME-RE structures the requirements gathering and elicitation process in four concurrent streams: decision objectives; data feasibility; analytical feasibility; and process-governance constraints, including regulatory constraints. When conflicts arise between streams, the methodology provides for controlled recursive refinement of those streams to resolve the conflicts. The results of the methodology have been validated using two sources: an expert survey with 25 practitioners and prior empirical research on how to structure unstructured data for evaluating risk. In summary, a conceptual architecture is introduced illustrating how non-functional requirements (e.g. stability and interpretability) influence system design and how unstructured information can be integrated as contextual signals for the purpose of informing risk evaluations while preserving core risk assessment capabilities. The results show that parallel and recursive requirements elicitation enables the development of credible, transparent and compliant ML systems.