Supervising Sentiment Models: Market Signals or Human Expertise?

We build a framework to examine how the training regime-rather than model architecture-drives the performance of financial sentiment models. Using firm-level news and parsimonious classifiers, we compare three supervision regimes: human-only, hybrid, and market-only (fully automated). The framework opens the "black box" of sentiment modeling by tracing how supervision shapes each component of the classifier. Across extensive tests, the hybrid regime consistently outperforms fully automated training in explaining variation in stock returns and trading volume, enhancing interpretability and economic relevance. Human input improves sentiment inference, offering new insights into information processing and price formation in financial markets.