FinDebate: Multi-Agent Financial Analysis

• FinDebate is a multi-agent, Retrieval-Augmented Generation framework that converts lengthy earnings call transcripts into coherent, institutional-style investment reports.
• It integrates domain-specific retrieval with five specialized analyst agents and a safe collaborative debate mechanism to produce evidence-grounded and multi-dimensional financial insights.
• The system delivers explicit long/short recommendations over 1-day, 1-week, and 1-month horizons while mitigating overconfidence and ensuring stance stability.

FinDebate is a multi-agent, RAG-based framework for financial analysis that generates institutional-style investment reports from earnings call transcripts by combining domain-specific retrieval, five specialized analyst agents, a report synthesis stage, and a single-round safe collaborative debate mechanism. Its stated targets are evidence-grounded analysis, multi-dimensional coverage of earnings, market reaction, sentiment, valuation, and risk, stance stability across revisions, and reduced overconfidence, with explicit long/short recommendations over 1-day, 1-week, and 1-month horizons (Cai et al., 22 Sep 2025).

1. Concept and analytical objective

FinDebate addresses a specific problem: converting long, unstructured earnings call materials into coherent analyst-style reports with explicit directional recommendations. The framework is motivated by two limitations identified in the source material. First, single LLMs can summarize earnings calls but tend to hallucinate facts, flip stance with small prompt changes, and underweight risk and sentiment. Second, traditional quant and fundamental approaches have numerical rigor but do not naturally ingest long, unstructured texts such as calls and Q&A, and therefore still require substantial analyst interpretation to produce tradeable views (Cai et al., 22 Sep 2025).

The framework is defined around four output properties. Reports are intended to be evidence-grounded, meaning traceable to earnings call content; multi-dimensional, meaning they cover earnings, market reaction, sentiment, valuation, and risk; stance-stable, meaning recommendations do not change arbitrarily between sections or revisions; and less overconfident and more reliable. The output format includes explicit long/short recommendations across three horizons: 1-day, 1-week, and 1-month (Cai et al., 22 Sep 2025).

This design places FinDebate within the broader class of debate-based reasoning systems, but with a constrained objective. It is not an open-ended adversarial debate to determine a stance from scratch. Instead, it is a refinement architecture in which a draft report is first synthesized and then revised under explicit safety constraints.

2. System architecture and specialized analyst agents

The architecture has three main modules: a Domain-Specific RAG Module, a Multi-Agent Analysis Module, and a Safe Collaborative Debate Mechanism. The pipeline is: input earnings call, then segmentation and indexing, then agent-specific retrieval and section generation, then report synthesis, and finally safe debate producing a final report $R^*$ (Cai et al., 22 Sep 2025).

The Domain-Specific RAG Module segments long earnings calls and related documents into semantically coherent chunks, encodes them with a finance-specific embedding model, and exposes a retrieval interface to all downstream agents. The Multi-Agent Analysis Module runs five specialized agents in parallel. Each agent receives a system prompt specifying role and credentials, a user prompt specifying task and structure, retrieves evidence relevant to its analytical dimension, and produces a structured section. The Safe Collaborative Debate Mechanism then refines the synthesized draft through Trust, Skeptic, and Leader roles while preserving the report’s directional calls (Cai et al., 22 Sep 2025).

Agent	Primary role	Core focus
Professional Earnings Analyst	Fundamental financial statement analysis	Revenue, EPS, margins, guidance, sector metrics
Professional Market Predictor	Multi-horizon market reaction forecasting	1-day, 1-week, 1-month reactions from call content
Professional Sentiment Analyst	Behavioral finance and credibility analysis	Tone, transparency, confidence, defensiveness
Professional Valuation Analyst	Fundamental valuation and fair-value judgment	DCF-style reasoning, catalysts, directional value view
Professional Risk Analyst	Comprehensive risk assessment	Credit, market, liquidity, operational, regulatory, idiosyncratic risk

The Professional Earnings Analyst uses the transcript plus retrieval focused on revenues, EPS, margins, guidance, and sector-specific metrics. For banks and financials, the stated focus includes net interest margin, asset quality, and capital ratios. The Professional Market Predictor is constrained to infer 1-day, 1-week, and 1-month reactions from call content alone, and is explicitly prohibited from using real-time prices or non-call data. The Professional Sentiment Analyst evaluates management confidence, caution, transparency, and credibility, especially in Q&A. The Professional Valuation Analyst performs qualitative DCF-style reasoning using growth, margins, capital allocation, and guidance, while avoiding external numeric modeling. The Professional Risk Analyst identifies credit, market, liquidity, operational, regulatory, and idiosyncratic risks and links them explicitly to management’s own discussion (Cai et al., 22 Sep 2025).

All five agents are instructed to maintain realistic confidence, typically 70–80%, and to base statements strictly on call content and RAG evidence. The Report Synthesis Agent then merges these five perspectives into a unified report and emits explicit recommendations for each horizon $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$, including position, conviction, catalysts, and risk-return commentary. No explicit aggregation formula such as $\sum_i w_i S_i$ is used; synthesis is performed by LLM reasoning guided by a structured template (Cai et al., 22 Sep 2025).

3. Retrieval layer, evidence grounding, and knowledge organization

FinDebate’s retrieval layer is domain-specific in both indexing and query design. Its corpus is primarily earnings call transcripts, specifically the ECTSum subset plus additional professional reports, and the framework is described as designed to scale to hundreds-page documents such as SEC filings and presentations. The segmentation strategy is a contextual chunking hierarchy: preserve paragraph boundaries first, then sentence boundaries, and only then split by lexical or token counts if necessary. This late chunking procedure is intended to preserve semantic coherence and avoid contextless fragments (Cai et al., 22 Sep 2025).

The vector database is ChromaDB, and embeddings are produced by FinLang, described as a finance-tuned embedding model derived from BGE. The retrieval formulation is given as

$$\mathbf{v}_i = f_{\text{FinLang}}(\text{segment}_i)$$

with similarity typically implemented as cosine similarity: $$\text{sim}(\mathbf{q}, \mathbf{v}_i) = \frac{\mathbf{q} \cdot \mathbf{v}_i}{\|\mathbf{q}\| \|\mathbf{v}_i\|}$$ where $\mathbf{q}$ is the embedding of the agent’s query and $\mathbf{v}_i$ is the embedding of segment $i$. Top-$k$ segments are retrieved for downstream prompting (Cai et al., 22 Sep 2025).

Retrieval is multi-level and task-conditioned. Four retrieval dimensions are specified: general financial performance; specialized financial metrics; market sentiment and risk; and multi-query integration. Agent queries are aligned to these dimensions. The earnings agent uses queries such as “Revenue, earnings, beat/miss, guidance, forward outlook” and, for financial institutions, “Net Interest Margin, ROA, ROE, Non-Performing Assets, provision for loan losses.” The market agent emphasizes beat or miss, earnings surprise, guidance surprises, positive or negative catalysts, and management confidence. The sentiment and risk agents query management confidence, tone, headwinds, analyst concerns, risk management, credit risk, and uncertainties (Cai et al., 22 Sep 2025).

Grounding is operationalized through prompt constraints rather than an explicit attribution loss. Agents are repeatedly instructed to base all assessments on verifiable information from the actual earnings call, to include quotes, and to avoid hypothetical scenarios or external data. This suggests that FinDebate treats retrieval not as a generic context augmenter but as a control layer for factual discipline in long-form financial reasoning.

4. Safe collaborative debate protocol

The safe debate mechanism is the architecture’s distinctive design contribution. It uses three debate agents—Trust, Skeptic, and Leader—and differs from open-ended multi-round debate by enforcing stance preservation and bounded refinement. The Trust Agent strengthens the existing report by preserving all long/short recommendations and conviction levels while adding evidence and improving clarity. The Skeptic Agent adds vulnerabilities, risk analysis, hedging, and scenario discussion, again without changing recommendations or conviction levels. The Leader Agent synthesizes these revisions into the final report while preserving catalysts, timelines, and structural elements (Cai et al., 22 Sep 2025).

Algorithm 1, “Safe Collaborative Debate,” takes as input the original synthesized report $R_0$ and the bundle $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$0 of the five specialized-agent outputs, and returns the optimized report $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$1 and a debate log $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$2. The stages are: $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$3

$$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$4

$$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$5

A safety check first validates structure; if has_recommendations(R_0) is false, the system returns $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$6 unchanged. A final check tests core_compromised(R^*, R_0); if stance or horizons changed, the system reverts to $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$7 (Cai et al., 22 Sep 2025).

The protocol is explicitly single-round. The reason given is avoidance of thematic drift that multi-round debate can cause. Overconfidence mitigation is procedural rather than metric-based. All agents are instructed to use 70–80% conviction rather than 95–99%, to mention uncertainty and risk, and to avoid near-certain language. Trust is constrained to strengthen evidence rather than conviction; Skeptic inserts risk and mitigation; Leader preserves existing conviction labels and directional recommendations. No numeric calibration loss such as a Brier score is defined (Cai et al., 22 Sep 2025).

This makes the debate stage a post-synthesis quality-control mechanism rather than a consensus engine. A plausible implication is that FinDebate uses debate primarily to densify evidence and broaden risk articulation while suppressing the instability that can arise when multiple agents are allowed to renegotiate the thesis itself.

5. Decision synthesis, evaluation protocol, and reported results

The report synthesis stage produces multi-horizon recommendations with a fixed schema. For each horizon $$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$8, FinDebate outputs

$$h \in \{\text{1-day}, \text{1-week}, \text{1-month}\}$$9

and

$\sum_i w_i S_i$0

The template explicitly includes position, conviction, key catalyst, and horizon-specific rationale. The paper notes that one can conceptually think of an unformalized ensemble,

$\sum_i w_i S_i$1

followed by

$\sum_i w_i S_i$2

but no explicit numeric aggregation is provided (Cai et al., 22 Sep 2025).

Evaluation is conducted on Earnings2Insights, using 40 earnings call transcripts from ECTSum and 24 professional analyst reports. Five base LLMs are used: GPT-4o, Gemini 2.5 Flash, Llama 4 Maverick, DeepSeek-R1, and Claude Sonnet 4. Uniform generation settings are temperature 0.6, max output length 6500 tokens, top-$\sum_i w_i S_i$3 0.85, and frequency penalty 0.1. Baselines are zero-shot, standard RAG with general-purpose embeddings, and multi-agent without debate; FinDebate is domain-specific RAG plus multi-agent plus safe debate (Cai et al., 22 Sep 2025).

The LLM-based evaluation follows Goldsack et al. and scores two major dimensions: Textual Quality, comprising readability, language abstractness, and coherence; and Financial Analysis Professionalism, comprising financial key point coverage, background context adequacy, management sentiment conveyance, future outlook analysis, and factual accuracy. Reports are scored on a 1–4 scale. Human evaluation measures whether experts can make correct long/short decisions for next day, next week, and next month, and also rates clarity, logic, persuasiveness, readability, and usefulness (Cai et al., 22 Sep 2025).

The reported results show gains across all five base LLMs. For GPT-4o, the average evaluation score rises from 2.97 for zero-shot to 3.21 for standard RAG, 3.39 for multi-agent without debate, and 3.58 for FinDebate. Across models, gains from zero-shot to FinDebate are reported as +0.59 – +0.62, about 20.4% relative, and all improvements are statistically significant under paired $\sum_i w_i S_i$4-tests with $\sum_i w_i S_i$5. The contribution of safe debate relative to multi-agent without debate is reported as an additional 0.16–0.19. Human evaluations likewise show improvements in day, week, and month decision accuracy and in clarity, logic, persuasiveness, and usefulness. The paper explicitly states that it does not compute live trading metrics such as Sharpe ratio because no live trading was executed (Cai et al., 22 Sep 2025).

6. Relation to debate research, limitations, and future directions

FinDebate is situated against broader debate research by design rather than by token-budget optimization. In a matched-ceiling study of open-weight 8B models, greedy decoding, three-sample voting, and a two-agent critique-revise debate were compared under a fixed ceiling of 960 generated tokens per example. That study found that the best fixed protocol depends on model and dataset, that an oracle router on MuSiQue has large headroom, and that vote entropy predicts where debate is safe rather than where debate is needed. It also found that simple vote-entropy thresholding directionally beats the best fixed protocol on both tested models, but only by about 1–2 percentage points and without statistical significance, while many debate-helpful cases occur when voting is unanimous but wrong (Hu et al., 10 May 2026). This suggests that if FinDebate were extended to token-budget-aware routing among direct answering, voting, and debate, cheap disagreement signals would likely recover only a small fraction of potential headroom.

A different empirical line of work reports that homogeneous unguided multi-agent debate with 7–8B models can exhibit sycophantic conformity, contextual fragility, and consensus collapse, while consuming 2.1–3.4$\sum_i w_i S_i$6 more tokens than isolated self-correction for equal or lower accuracy (Bertalanič et al., 29 Apr 2026). FinDebate differs structurally from that setup in three explicit ways already present in its design: agents are domain-specialized rather than homogeneous, the debate is single-round, and stance preservation is enforced. This suggests that its Trust–Skeptic–Leader mechanism is better interpreted as constrained report revision than as unconstrained consensus formation.

There are also evaluation-related caveats. LLM-based debate evaluators have documented positional bias, lexical bias, order bias, and end-of-discussion bias, including a tendency to favor the second candidate response and the side that concludes the debate (Liu et al., 2024). Because FinDebate uses GPT-4o as an evaluator in one part of its experimental protocol, this literature is relevant to interpretation of automatic scoring, even though the framework also includes human evaluation.

The limitations stated for FinDebate itself are narrower and application-specific. Experiments are limited to earnings calls in Earnings2Insights; there is no live trading evaluation; retrieval quality depends on FinLang and chunking; performance may vary with base LLM quality; and multi-agent RAG plus long reports plus debate is computationally heavier than a single-call summarizer. Real-world risks include over-reliance on automatically generated investment recommendations, reproduction of management-crafted narratives rather than underlying reality, regulatory concerns around automated recommendations, and the possibility that stance preservation may block legitimate course corrections when the initial thesis is wrong (Cai et al., 22 Sep 2025).

The proposed future directions are extension to broader financial tasks, dynamic confidence adjustment mechanisms, integration of real-time market data, and possible downstream use in execution systems or portfolio construction pipelines. A plausible implication is that future versions of FinDebate would combine the current stance-stable, evidence-grounded reporting framework with more explicit calibration and market-state awareness, while preserving the architecture’s central emphasis on bounded refinement rather than unconstrained argumentative drift.