THE ECONOMIC COST OF FINANCIAL FAKE NEWS IN EUROPEAN CAPITAL MARKETS

Abstract

Financial news misinformation is a systemic risk that distorts price discovery and capital allocation by widening information asymmetry and weakening investor trust. This study develops a scalable NLP and machine-learning framework to detect deceptive financial narratives in large digital corpora through Big Data. After text normalization, lemmatization, and stopword elimination, the framework contrasts TF-IDF with Word2Vec embeddings and trains Logistic Regression, Random Forest, and Gradient Boosting classifiers. Performance is assessed with Accuracy, Precision, Recall, F1-score, and ROC-AUC. Across models, TF-IDF provides stronger discrimination than Word2Vec; the TF-IDF Random Forest reaches near-perfect results (ROC-AUC 0.9999; Precision 0.9977). The emphasis on transparent, feature-based models supports auditability (for example, via feature importance) and helps limit harmful false positives that could suppress legitimate signals. The results indicate that high-precision, interpretable pipelines can reduce the verification gap in fast-moving information environments, mitigate macroeconomic costs of deceptive narratives, and inform DSA- and ESMAaligned market surveillance workflows. The framework is designed for deployment on streaming news feeds and large-scale platform archives.

JEL: G14, D82, C55.