ARGUS Survey of AI Driven Personal Finance Analytics and Budget Forecasting

Abstract

Personal-finance software is useful only when it can move from records to decisions without hiding uncertainty. We reviewed 25 papers published in 2025 and 2026 and grouped their evidence into six connected activities: UNDERSTAND, ANTICIPATE, PERSONALIZE, ADVISE, EXPLAIN, and TRUST. For every paper, we recorded the study design, data, method, reported result, limitation, and relevance to ARGUS. The stated review window runs from 1 January 2025 to 30 September 2026, while the locked-corpus audit was completed on 22 September 2026. The study set contains 11 observational surveys, six human experiments, four audits of model output, two computational studies, and two reviews. This distribution matters because only P02 directly evaluates both personal transaction classification and budget prediction. The remaining evidence is strongest on advice, adoption, privacy, vulnerability, explanation, and trust. We therefore do not treat an adoption coefficient or a favorable user rating as proof that a budgeting model works. We also compare the literature with the inspected ARGUS implementation. The current project uses deterministic categorization, exact-money arithmetic, budget comparisons, and answers grounded in the user ledger; it does not contain a trained forecasting model. Archived project records report 205 passing automated tests, acceptable grounded responses for ten controlled questions, and correct amount and currency extraction from two clear synthetic receipts. These are engineering checks on specific cases, not estimates of population accuracy or financial benefit. The review supports an auditable design for personal-finance assistance, but it also shows that learned forecasting, personalization, and long-term household outcomes still require direct evaluation.