FinSight: A Survey of Multi-Agent and Language-Model Approaches for Financial Intelligence

Financial analysis draws on filings, earnings calls, news, and market data, but the tools that analyse these sources are separate, so analysts end up reconciling sentiment, event, fundamental, risk, and forecasting signals by hand. This survey reviews ten core studies published between 2019 and 2025, along with 32 supporting works on financial language models, event extraction, sentiment-driven forecasting, risk assessment, multi-criteria company ranking, and multi-agent LLM systems. We propose a five-dimensional taxonomy and compare the core studies on eight analytical capabilities. Most of them handle one or two tasks in isolation, and in our reading none uses confidence scoring to reconcile conflicting outputs from different components. We identify six research gaps and describe how FinSight, a metacognitive multi-agent platform with a Confidence & Safety Checker, addresses each one through its architecture, data design, requirements, and evaluation plan. The paper ends with applications, open challenges, best practices, and a roadmap for future work.