FARE: Financial Agentic Reasoning and Evaluation for Earnings Call Transcripts

Abstract

Earnings call transcripts contain crucial information about the financial performance of companies across various industries. However, due to their unstructured nature, an automated interpretation system is challenging to build. This study presents Financial Agentic Reasoning and Evaluation (FARE) framework that analyzes earnings call transcripts and answers users' queries using an integration of retrieval-augmented generation (RAG) and multi-agent reasoning. The framework built using LangGraph contains six agents: Statement Extraction, Tone, Investment, Comparison, Tone Shift, and Simulation. Each agent performs a specialized task ranging from providing investment guidance to analyzing company sentiment. Due to the lack of evaluation metrics for assessing performance of agentic systems with earnings call transcripts, we introduced evaluation metrics to evaluate each agent using an LLM-as-a-judge system. The FARE framework was evaluated on a curated dataset of 168 quarterly earnings call transcripts from 21 companies. We compared the performance of four LLM backbones: Gemini 2.5 Flash, Gemini 2.5 Flash-Lite, Qwen2.5-7B-Instruct, and Llama-3.2-3B-Instruct which were integrated in the system. The results show that Qwen2.5-7B-Instruct achieves the strongest overall reasoning performance across multiple agents, while Gemini 2.5 Flash-Lite consistently demonstrates significantly lower latency and higher efficiency.