Neural Modernization for Automated Reconstruction of Legacy Financial Application Semantics

Abstract

The legacy financial applications, often built in COBOL and other third generation programming languages, are used to support the backbone of banking, insurance and payment processing systems throughout the financial world. The ability to reverse engineer the business logic and computation that are hidden inside these systems is essential for modernization, risk assessment and regulatory compliance, but manual reverse engineering is too time consuming and error-prone. This review aims at summarizing the advances that were made over the last four decades (1984-2022) in reverse engineering, and particularly in semantic reconstruction, from a rule-based to neural and transformer-based approach. Based on a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol, five converging research threads were identified through screening and thematic synthesis: program comprehension theory, distributed code representation learning, sequence-based neural approaches, transformer pretraining for source code, and empirical migration case studies in banking environments. The reported results show that neural code representation models like code2vec and CodeBERT are able to retrieve functions with an 80+ percent accuracy on curated benchmarks, and that documented COBOL-to-Java migration projects achieve 30-40 percent productivity improvements compared to complete re-implementation, with only two to four defects per thousand lines of migrated code. Nearly forty-five percent of program patches are accepted using automated program repair techniques based on neural machine translation that leverage defect corpora. The review finds that neural modernization pipelines work best when combined with, not replaced by, traditional slicing, control flow analysis and business rules extraction techniques, and that there are no large, labeled corpora of COBOL code, there are regulatory opacity requirements, and there is a high cost of validation that are the main challenges to industrial use.