• The UIL architecture addresses the limitations of unimodal models by combining a natural language processing (NLP) pipeline for sentiment extraction with a Long Short-Term Memory (LSTM)/Gated Recurrent Unit (GRU) price-action encoder.3
• It utilises a learned gating-based late fusion layer to dynamically weight the importance of sentiment versus price data based on current market states.4
• The system is trained with an asymmetric loss function that prioritises the identification of genuine crises over avoiding false alarms.5
• Validation through historical backtesting on events like the SVB collapse (2023), the UK Gilt crisis (2022), and the COVID-19 market crash (2020) demonstrated superior performance and earlier lead times in crisis detection compared to unimodal baselines.

‍