Multilingual Textual Signals and Realized Volatility Forecasting: A Hybrid Transformer-HAR-X Approach Applied to the S&P 500 and the MASI

Abstract

Realized-volatility forecasting supports risk management and allocation. GARCH and HAR-RV mostly exploit numerical memory, whereas LSTM-GARCH and FinBERT-LSTM hybrids are often monolingual and rarely receive out-of-sample economic evaluation. We propose a bilingual FinBERT-CamemBERT framework that constructs the daily textual stress index Sₜ and the event vector Eₜ, which enter HAR-X for the S&P 500 and the MASI.

The corpus contains 12.3 million French and English items over 2015-2024. Expanding-window validation reserves 2023-2024 for final testing. Against HAR-RV, HAR-X reduces the S&P 500 RMSE by 21.1%; the Diebold-Mariano test yields DM = -2.44, p = .015, and qBH = .033. For the MASI, DM = -2.62 and qBH = .033. Benjamini-Hochberg correction covers six comparisons. The net-of-cost Sharpe ratio rises from .64 to .81, but this gain remains descriptive without a direct Ledoit-Wolf test. Twenty-three of twenty-four monthly differentials are positive, with p < .001 under an exact sign test.

The framework connects multilingual textual information, heterogeneous memory, and allocation. SHAP assigns a major predictive role to Sₜ and Eₜ without establishing a causal relationship. Limitations concern annotation reliability, foreign exchange, and concept drift.

JEL Classification: C53, C58, G17, G11.

Paper type: Empirical Research.