Factor-Based Quantile Forecasting With Textual Data

Abstract

Words matter for predicting tail risks. We propose an attention mechanism embedded in a quantile factor model, yielding information that is quantile-specific, target-specific, and horizon-specific. We establish new asymptotic results and show empirically that targeted textual data improve quantile forecasts of exchange-rate returns and industrial production growth relative to strong benchmarks and other conditional-quantile models. Bigrams and trigrams drive these gains, extending evidence that collocations enhance forecasting. A portfolio-management application yields better allocations and higher risk-adjusted performance. Robustness checks include a synthetic-misinformation stress test, which shows that fabricated news does not create spurious predictability. Results are consistent across targets, horizons, and both rolling and recursive schemes.