The Analyst in the Prompt: Role, Retrieval, and Memory Biases in LLM Financial Analysis
Ahmed Asaad, Amr Mohamed, Yang Zhang, Omneya Abdelsalam
cs.CL · cs.CE · q-fin.PM
Quick verdict
A careful 12-LLM audit on 3,575 SEC filings finds role prompts shift supposedly neutral evidence scores; convincing, with causal and external-validity gaps.
Summary
This paper audits whether personalization context — investor personas set as assistant roles, or user memory profiles — spills over into an LLM output field explicitly defined as a neutral, evidence-based judgment. Across 3,575 SEC 10-K/10-Q filings from S&P 500 firms (2013–2024) and twelve open-weight LLMs from five families, the authors compare persona-conditioned retrieval (PR), neutral retrieval with byte-identical evidence across personas (NR), and a memory condition (MEM) that hands the same mindset to a neutral assistant as third-person user context. Filing fixed-effects regressions show that about 69% of the role effect on the evidence score survives identical retrieval on average (44–89% per model), so interpretive drift, not retrieval drift, dominates. Replacing the assistant role with a user profile shrinks the effect (median 73% for the strongest ShortSeller pair; 51–87% across models), and a dual-output schema keeps leakage below 0.5 for all models (0.05–0.37) without eliminating it, independently of scale. Scores retain within-persona ranking signal (top-minus-bottom quintile abnormal return about +2.96% panel mean), and residual spillover is 2–4× larger on small-cap filings. Limitations include one retrieval encoder and prompt family, an accounting rather than causal-mediation decomposition, possible pretraining memorization, and mechanism checks on only two models.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Most user-context spillover into the evidence-based score comes from interpreting identical evidence differently under different roles (69% panel-mean retention under byte-identical retrieval), not from retrieving different evidence.
- The same investor mindset causes substantially less spillover when framed as a user memory profile than as the assistant's role (median 73% shrinkage for the ShortSeller/DownsideProtection pair).
- A dual-output schema reduces but never eliminates leakage (all twelve models below 0.5, range 0.05–0.37), with effectiveness that is model-specific rather than scale-driven.
- Role conditioning shifts score levels but within-persona rankings retain forward-looking signal (panel-mean top-minus-bottom quintile abnormal return ≈ +2.96% at one year).
- Residual spillover is 2–4× larger on small-cap filings, indicating invariance failures concentrate on less standardized documents.
Strengths
Clean three-condition isolation design
PR→NR pins retrieval to byte-identical chunk sets (verified: median Jaccard 1.000 under NR vs ~0.66 under PR, App. D.3), separating evidence selection from interpretation; MEM then varies only the prompt placement of the same mindset.
Scale, cross-model breadth, and verbatim artifacts
3,575 filings × 12 models across 5 families; Appendix C reproduces the system prompt, persona/profile blurbs, retrieval queries, and JSON schema verbatim; strong-persona cells report |t|>20 on every model (App. D.2).
Multi-layer triangulation of one mechanism
Robustness beyond the headline
Honest invariance framing
Weaknesses
Accounting decomposition, not causal mediation
Neutral reference conflates invariance with validity
Single encoder, single prompt family, uncounterbalanced blurbs
Downstream return inference is under-powered as reported
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a sharp invariance question: when an LLM output field is explicitly defined as a neutral, evidence-based judgment (an EvidenceScore in [−1, +1] over retrieved filing chunks), does user context — investor personas installed as the assistant's role, or the same mindsets delivered as user memory profiles — move it? The setting is 3,575 SEC 10-K/10-Q filings from S&P 500 constituents (2013–2024, at most six filings per firm) and twelve open-weight models from five families (DeepSeek, gpt-oss, Llama, Qwen, Gemma). Three matched conditions separate candidate channels: Persona-Retrieval (PR: the persona conditions both query and prompt), Neutral Retrieval (NR: every persona reads byte-identical top-8 chunks), and MEM (a neutral assistant receives the mindset as a third-person user profile). The four central claims: (1) about 69% of the PR role effect on the evidence score survives identical retrieval (44–89% per model; ShortSeller panel median 89.1%), so interpretive drift dominates retrieval drift; (2) role→profile reframing shrinks the effect (ShortSeller/DownsideProtection: 51–87%, median 73%); (3) a dual-output schema keeps leakage below 0.5 for all twelve models (0.05–0.37) without eliminating it, with no relationship to scale; (4) within-persona score rankings retain forward return signal (panel-mean top-minus-bottom quintile ≈ +2.96% at one year), making the invariance failure practically consequential.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'user-context spillover' in this paper?
It is the movement of the LLM's evidence-based score — a field instructed to be a neutral reading of the filing, independent of the user — when user context such as an investor persona/role or a user memory profile changes. The paper measures how far that context leaks into the supposedly invariant field across 3,575 SEC filings and twelve LLMs, finding on average 69% of the role effect survives even when every persona reads identical retrieved evidence.
Which LLMs showed the most role bias in financial analysis?
By mean role bias on the evidence score, Llama-3.1-8B-Instruct was largest (0.40) and gemma-4-31B-it smallest (0.02), with a panel mean of 0.20 across twelve models. Under the dual-output schema, leakage ranged from 0.05 (Llama-3.3-70B-Instruct) to 0.37 (Qwen3.6-35B-A3B), and bias did not decrease monotonically with model size.
How can personalization bias in LLM financial advice be reduced?
The paper tests two mitigations: presenting the investor mindset as a third-person user profile instead of an assistant role (shrinking the effect by a median 73% for the strongest persona pair), and a dual-output schema that separates an evidence-based score from a personalized score (leakage below 0.5 for all twelve models). Neither removes the bias completely, and effectiveness varies substantially by model, so both should be treated as risk reduction, not guarantees.
Did the biased LLM scores actually predict stock returns?
Within a fixed persona, rankings by the evidence-based score retained signal: filings in the top score quintile outperformed the bottom quintile by about +2.96% on average over one year (market-adjusted, panel mean across twelve models), with eight of nine personas positive and RiskManager roughly flat. The bias shifts score levels across personas rather than destroying within-persona ranking information.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All numbers cited were taken from prose or tables in the extraction; figure-internal values (e.g., Figures 2–3 axis data) appear as interleaved text and were not relied upon.
- The extraction ends cleanly after the Appendix D.17/Table 10 discussion; no truncation was observed, though Table 4 column alignment was read from the panel (a)/(b) captions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.