Low-Latency Multimodal Risk Engine: Real-Time Financial Risk Assessment using Streaming Text and High-Frequency Time Series
Vinitta Sunish1, Lydia Suganya2, Abhilasha Patil3, Siddhi Ambre4, Soumyamol P.S5
Assistant Professor
1Computer Engineering Department of 1st Author,
1Thakur College of Engineering and Technology of 1st Author, Mumbai, India
vinitta.sunish@thakureducation.org,lydia.suganya@thakureducation.org, abhilasha.patil@thakureducation.org, siddhi.patade@thakureducation.org , soumyamol.ps@thakureducation.org
Abstract - Traditional risk models fail during extreme market events. They ignore the qualitative clues in corporate narratives, news and investor discussions. New multimodal Large Language Model (LLM) frameworks have demonstrated superior predictive accuracy. This is done by combining textual and quantitative data. Their inference latency (seconds) prevents deployment in real-time monitoring systems. This research solves this critical gap by proposing a new low-latency multimodal architecture. This is engineered for financial risk assessment in sub-second. This system uses lightweight FinBERT-class encoders for text series. It uses high-frequency transformers for numerical data. It combines this data using an efficient cross-modal attention fusion mechanism. Tested the system on Chinese A-share companies (2001–2024). It is also augmented with real-time news streams. The proposed system achieves less than 100ms speed or end-to-end inference. It preserves the predictive power of heavyweight batch-oriented LLMs. The main contributions are (i) the first real-time risk engine to combine the process of textual data-unstructured and time-series data-high-frequency at latency of sub-second, (ii) a lightweight fusion strategy boosts the performance gap between LLMs and edge-deployable systems, and finally, (iii) empirical validation shows significant early-warning gains during market stress.
Key Words: multimodal learning, financial risk, deep learning, time-series, text analysis, real-time prediction
