Consumer Complaint Mining Through Topic Modeling and Sentiment Analysis: A Proof-of-Concept Analytical Framework for Decision Support
TL;DR
This study presents a proof-of-concept analytical framework for multi-dimensional complaint analysis and uncertainty-aware evidence synthesis, with potential use in human-in-the-loop decision support, and applies it to 99,434 CFPB complaint records spanning from March 2015 to March 2026.
Abstract
Consumer complaint narratives are an underused source of business and regulatory evidence. This study presents a proof-of-concept analytical framework for multi-dimensional complaint analysis and uncertainty-aware evidence synthesis, with potential use in human-in-the-loop decision support, and applies it to 99,434 Consumer Financial Protection Bureau (CFPB) complaint records spanning from March 2015 to March 2026. The empirical workflow integrates BERTopic topic discovery, two pretrained sentiment checkpoints (ProsusAI/finbert and cardiffnlp/twitter-roberta-base-sentiment-latest), exploratory clustering of complaint records, and descriptive temporal summaries. A conceptual stock–flow and feedback representation links observed complaints, analytical alerts, response capacity, and unresolved issues; these relations are not causally estimated or simulated. BERTopic produced 89 non-outlier topics. A probability-based outlier-reassignment stage assigned 80.46% of records to these topics, while 19.54% remained unassigned and were retained as an uncertainty queue. A FinBERT-prediction-stratified benchmark of 600 complaint records was independently coded by two annotators (raw agreement = 87.5%; Cohen’s κ = 0.758). Against Annotator 1 as the prespecified reference, Cardiff RoBERTa aligned better with the labels on the prediction-stratified benchmark (κ = 0.395; sample accuracy = 67.3%) than FinBERT (κ = 0.195; sample accuracy = 46.3%), although neither model supports autonomous use. The four-group K-means solution is reported as exploratory because the highest observed Silhouette coefficient was only 0.106 and favored K = 2. Temporal patterns coincided with selected external events, but no causal effect is claimed. The contribution is therefore a transparent, uncertainty-aware analytical architecture whose potential organizational uses require human review and operational validation.