Let the Laser Beam Connect the Dots: Forecasting and Narrating Stock Market Volatility.

This article focuses on forecasting market volatility using news events and enhancing model interpretability through narrative generation. It introduces LASER (Long- and Short-term memory Retrieval), a deep learning architecture that combines pretrained BERT for encoding news titles and recurrent independent mechanisms (RIMs) for modeling temporal dynamics across flexible short- and long-term memory horizons. To address the interpretability challenge of deep neural networks, the authors propose BEAM, a controlled text-generation algorithm leveraging the pretrained GPT-2 language model to produce human-readable narratives that explain the evidence behind volatility predictions. Empirical results on Wall Street Journal news and S&P 500 data demonstrate that LASER outperforms baseline models in predicting high-volatility incidents, while BEAM generates more fluent and informative explanations compared to existing decoding methods. The LASER-BEAM pipeline is presented as a generalizable framework for sequential text-based forecasting tasks with interpretability needs.