A Human Factors Framework for Operational Risk Management in Banking Using Deep Learning and Large Language Models

Operational risk management (ORM) in financial institutions has traditionally relied on quantitative loss event databases that capture what went wrong, but rarely why from a human factors perspective. This paper proposes a comprehensive Human Factors Framework for ORM that integrates the Human Factors Analysis and Classification System (HFACS) taxonomy with deep learning and large language models (LLMs) to extract, classify, and predict human-factor-driven operational risks from unstructured consumer complaint narratives. Using the U.S. Consumer Financial Protection Bureau (CFPB) Consumer Complaints Database (300,000 banking narratives), we (1) define and validate an eight-factor HFACS-Banking taxonomy through a Delphi study with five domain experts (Cohen’s κ = 0.86), (2) compare three extraction approaches—regex, supervised BERT, and Mistral-7B zero-shot LLM—achieving 0.81 average F1 with the LLM-distilled BERT relative to Mistral-generated reference labels, (3) propose a hybrid HF-BERT-BiLSTM-Attention architecture that fuses contextual text embeddings with structured features and HFACS factor probabilities for predicting a four-class company-response-based complaint severity proxy, and (4) provide explainability through SHAP feature attribution applied to the Random Forest baseline and attention-weight visualization of the proposed neural model. Under a temporal split comprising training data from 2014 to 2022, validation data from 2023H1, and held-out test data from 2023H2 to 2025, the proposed model achieves 91.42% accuracy and 89.78% macro F1 (95% CI from 1000-iteration paired bootstrap: [89.34, 90.21]), outperforming Random Forest (+9.84% F1, p < 0.001 Bonferroni-corrected), BiLSTM (+5.46%, p < 0.001), FinBERT (+3.21%, p = 0.003), and BERT-only (+3.92%, p = 0.002) baselines. Under the secondary random-split ablation analysis, removing the HFACS features and replacing additive attention with mean pooling reduced macro-F1 by 3.79 and 2.35 points, respectively. The findings demonstrate the retrospective feasibility of integrating theory-grounded human-factor representations with neural language models for complaint-outcome analysis; prospective institutional validation is required before operational use.