A comparative typology of generations of machine learning models monolithic and hybrid for financial prediction a theoretical benchmark against traditional econometric models

Abstract

The simultaneous rise of deep architectures, financial Transformers, and hybrid models has fragmented the literature on financial time-series prediction into sub-corpora that are hardly comparable. No consensual typological framework currently classifies the successive generations of machine-learning models applied to finance. This study proposes a multi-level typology and a theoretical benchmark of five families (classical econometrics, shallow ML, deep learning, advanced AI and hybrids) along seven discriminant dimensions. The protocol mobilizes PCC under PRISMA-ScR on a stabilized corpus of 58 studies (1990–2026). Three formal propositions structure the contribution. P1 no generation is uniformly superior; P2 hybrids dominate conditionally under nonlinearity, regime breaks and low signal-to-noise; P3 financial-theory compatibility decreases (partial order) as representational capacity grows, except under theory-data hybridization. Beyond conceptual architecture, the framework delivers an operational decision grid for asset managers, banking model risk management and regulators (SR 11-7, TRIM, EU AI Act). The article fits Gregor’s (2006) typology (types II and IV) and provides a replicable instrument for the principled selection of models.

JEL Classification: C38, C45, C53, C58, C63, G17

Paper type: Theoretical Research