FUSION
2025 IEEE First International Conference on Innovations in Engineering and Next-Generation Technologies for Sustainability (ICINVENTS)
Aditya S. Chandra, Anisha S. Chandra, Varish Gahlot
10.1109/ICINVENTS64613.2025.11401395Varish built the fusion layer: the part of the system concerned with integrating structured financial data with narrative and other unstructured sources, so the language model can navigate between them.
- Tailored data ingestion
- Financial-specific text chunking
- Dual index: keyword and vector search
- Retrieval with re-ranking
- LLM-guided analysis
The research question
How can narrative financial information be fused with structured data so that a language model's analysis is both more relevant and clearer?
Market forecasts usually rest on structured data such as stock prices, company financials and economic trends. Analyst opinions, earnings-call transcripts and news add context that numbers alone can miss. FUSION introduces a system designed to integrate these unstructured insights into market forecasting.
Method
FUSION follows a multi-stage pipeline: tailored data ingestion, financial-specific text chunking, and a dual-indexing setup that combines keyword-based and vector-based search for more nuanced retrieval.
The authors tested four retrieval strategies (classic keyword search, pure vector retrieval, hybrid methods, and vector retrieval with re-ranking) on a curated dataset of 315 financial documents and 20 benchmark questions, cross-validated with FinQA and assessed using ROUGE scores and expert feedback.
- Four retrieval strategies tested: classic keyword search, pure vector retrieval, hybrid methods, and vector retrieval with re-ranking
- A curated dataset of 315 financial documents and 20 benchmark questions
- Cross-validation with the FinQA dataset
- Evaluation with ROUGE scores and expert feedback
Varish's contribution
FUSION is collaborative work by the same three co-authors. Varish built the fusion layer: the part of the system concerned with integrating structured financial data with narrative and other unstructured sources, so the language model can navigate between them.
Findings
According to the published abstract, vectorization combined with re-ranking significantly improved both the relevance and the clarity of the model's outputs. The authors present FUSION as a scalable way to bridge narrative financial analysis and structured data.
Limitations and future directions
The paper identifies practical challenges as areas for future work, including integrating real-time data feeds and adapting language models to financial language. As with SMART-Eq, results describe the curated evaluation dataset rather than deployment in live markets.