ML in finance: tools, limits, and real‑world use
Podcasts and recent coverage stress that tabular ML (XGBoost, LightGBM) still dominates production finance models, while NLP and deep learning are growing for sentiment tasks — but Stanford warns black‑box models introduce new systemic risks unless rigorously validated. The same week saw examples of off‑the‑shelf ML trading EAs and practical advice to focus on feature engineering, explainability, and production hardening. (gsb.stanford.edu) (yoforex.org) (youtube.com)
Stanford Graduate School of Business published “AI Could Predict the Next Financial Crisis — But There’s a Catch” on March 19, 2026, arguing that high‑dimensional predictive models can flag vulnerabilities but cannot by themselves establish causal mechanisms and may create moral hazard if used without policy context. (gsb.stanford.edu) U.S. bank supervisors still rely on Federal Reserve/OCC Supervisory Letter SR 11‑7 for model risk management and exam expectations, a framework examiners are applying to AI/ML systems despite it being issued in 2011. (federalreserve.gov) The UK’s FCA, PRA and Bank of England published coordinated AI approaches on April 22, 2024, and the PRA’s SS1/23 model risk statement took effect for in‑scope firms on May 17, 2024, explicitly extending scrutiny to AI and machine‑learning models. (hoganlovells.com) Academic and industry research continues to evaluate gradient‑boosted tree frameworks in finance: an IEEE study benchmarked XGBoost, LightGBM and CatBoost on a 50‑GB dataset of 2.3 million records and ~190 features for credit‑risk assessment. (ieeexplore.ieee.org) Specialist projects and marketplaces show off‑the‑shelf trading automation is widespread: YoForex lists an “AI Gold Trading EA V3.3 MT5” posted March 19, 2026 by developer madhuparnachaki_6262 priced at ~$145 with 41 page views, and the site catalogs hundreds of MT4/MT5 EAs aimed at retail and prop‑firm traders. (yoforex.net) Empirical work on explainability highlights risks and mitigation steps: a recent MDPI study quantified instability in SHAP attributions across repeated XGBoost credit‑risk model runs and recommended stability testing for regulatory compliance. (mdpi.com) MLOps research and industry guides stress production hardening practices now common in financial teams, including automated drift detection, CI/CD for models, model versioning, and retraining pipelines—practices shown to reduce deployment time and catch issues far faster in mature teams. (arxiv.org) Regulatory‑facing compliance toolkits and SR 11‑7 supplements explicitly list documented feature engineering, independent validation, auditable explainability outputs, and continuous monitoring as required controls for AI/ML models used in credit, fraud and trading workflows. (brianonai.com)
