Media Sentiment and Reputational Risk in a Public Mortgage Institution: A Daily Index of Media-Derived Reputational Pressure for INFONAVIT

Reputational risk is hard to quantify, and more so for public institutions in emerging markets. We build a daily index of media-derived reputational pressure for INFONAVIT, Mexico’s largest public housing lender, from 138,789 news articles published by 9495 outlets between November 2019 and July 2024. Transformer-based sentiment classification, neural topic modeling, emotion detection and time-series anomaly analysis are combined into a Reputational Risk Score (RRS), bounded between 0 and 100, that measures media pressure, not reputational losses. Agreement between the BERT classifier and the monitoring service’s human coders is negligible (Cohen’s kappa = −0.0019), so absolute sentiment levels cannot be read; replacing the classifier with two Spanish-language models leaves the index almost unchanged (r = 0.941 and 0.952). Rolling z-scores flag 63 crisis days in 53 episodes over 1223 days with at least 30 articles, and the highest peaks correspond to fraud investigations, operational changes affecting borrowers and incidents at institutional offices. The fraud-related detections survive all 48 detection settings, and the main detections are robust to alternative weights and coverage thresholds, but they require the tone components, and media volume carries 78% of the variance of the composite. Because episodes are interpreted ex post, we present an auditable record of media pressure rather than external validation.