Digitisation of banking in Southeast Asia has created a need for risk-assessment methods that work for thin-file customers and the underbanked, for whom traditional credit scoring is ineffective. This systematic literature review traces the shift from isolated machine-learning classifiers to hybrid spatiotemporal deep-learning architectures for financial-risk detection. Following PRISMA 2020, a search of Web of Science, Scopus and IEEE Xplore returned 58 eligible peer-reviewed records published between 2020 and 2025, comprising 56 primary empirical studies used for quantitative coding and two secondary reviews retained as methodological background. Publication volume rose from three studies in 2021 to 39 in 2025, indicating rapid expansion. Among the primary studies, 31 (55%) couple graph learning with cross-time-step recurrence and 25 (45%) do not, 17 of which encode time by other means. Coverage of the review questions is uneven, as every study proposes an architecture and 51 couple spatial with temporal modelling, yet only 28 address data engineering for input deficiency. Hybrid architectures can outperform isolated spatial or temporal models on dense benchmarks, but this advantage is contingent on sufficient graph density, behavioural sequences longer than a single transaction, and imbalance-aware evaluation, conditions that remain largely untested in sparse thin-file settings.