SMART-Eq
2025 World Skills Conference on Universal Data Analytics and Sciences (WorldSUAS)
10.1109/WorldSUAS66815.2025.11199272Structured data
Company filings, financial data
Unstructured text
Market commentary, narrative sources
Retrieval layer
My contribution
Large language model
Analysis drawing on multiple sources
The question
Equity forecasting has traditionally relied on structured data such as prices, fundamentals and macroeconomic indicators. Unstructured sources, including analyst reports, earnings-call transcripts and financial news, carry qualitative insight that numbers alone miss. SMART-Eq explores how the two can be combined so that a large language model can draw on multiple sources when analyzing financial information.
My contribution
I built the retrieval layer: the techniques that identify and connect relevant information across structured financial data and unstructured text, so that the model works from connected sources rather than a single one.
What the paper reports
According to the abstract on IEEE Xplore, vectorization combined with re-ranking improved both the precision of retrieved information and the contextual accuracy of the model's responses.
- A multi-stage pipeline: data ingestion, domain-specific text chunking and dual indexing (lexical and vector-based) for hybrid retrieval
- Four retrieval approaches compared: text search, vector retrieval, hybrid retrieval, and vector retrieval with re-ranking
- A custom dataset of 315 financial documents with 20 control questions, validated against the FinQA dataset
- Evaluated with ROUGE metrics and expert review