Fund2Persona: A Framework for Building and Refining Financial Advisor Personas from Fund Disclosure Data

Abstract: Demand for personalized financial advising is growing, but consistent advisor expertise is difficult to obtain, scale, and encode in LLM systems. Simple persona prompts rarely specify how a financial advisor should reason and often drift toward generic recommendations. We propose Fund2Persona, a framework that grounds financial-advisor personas in fund disclosures, holdings transitions, market context, and manager commentary, then refines them through an agentic actor--scorer--patcher loop. We evaluate the resulting personas on held-out holdings-transition reconstruction and manager-commentary alignment, where they better recover portfolio decisions and grounded manager interpretation than generic baselines. We further study two downstream diagnostics: market-scenario generation, where persona retrieval broadens plausible investment views beyond repeated generic rollouts, and advisory dialogues grounded in investor profiles, where matched personas give more specific and useful advice than a generic advisor. These results suggest that fund-data-grounded financial-advisor personas can make manager-specific investment expertise portable rather than merely changing an LLM's surface style.

• The paper introduces Fund2Persona, a three-stage framework that synthesizes and refines financial advisor personas using fund disclosures, holdings transitions, and market context.
• It leverages quantitative metrics like active-delta accuracy and manager commentary alignment to validate its portfolio reconstruction and narrative consistency.
• The framework enhances advisory realism by producing personas that faithfully replicate fund-specific decision logic over generic LLM-based approaches.

Fund2Persona: Synthesizing and Refining Financial Advisor Personas from Fund Data

Problem Statement and Motivation

The paper introduces Fund2Persona, a framework addressing the limitations of standard persona-prompting techniques for financial advisory roles in LLM systems (2606.29793). Traditional persona construction, based on surface-level role labels or limited disclosure-driven prompts, yields generic recommendations that fail to mimic the nuanced, policy-conditioned reasoning of real financial managers. Given the increasing integration of LLMs into user-facing advisory contexts, there is strong demand for advisor personas that reflect consistent, data-grounded investment philosophies, risk postures, and style-specific decision logic. The Fund2Persona framework exploits structured fund disclosure data, temporal holdings transitions, market context, and manager commentary to synthesize and iteratively refine LLM-based advisor personas with high fidelity to the original fund’s decision making.

Framework Architecture

The core of Fund2Persona consists of a three-stage agentic loop: actor, scorer, and patcher. The process begins with the construction of an initial persona based on the fund’s regulatory disclosure documents (D), initial holdings snapshot ($H_0$), and categorical fund attributes. This persona encapsulates explicit and implicit investment mandates, risk controls, core sector or style tilts, and turnover discipline.

Refinement proceeds via interpretive feedback using actual fund behavior over replay transitions ($H_{t-1} \rightarrow H_t$): the persona is tested on its ability to reconstruct end-of-period portfolio weights given market context and realized security-level returns. The active delta (i.e., residual after correcting for mechanical buy-and-hold drift) serves as the behavioral signal of manager action, and the reward is measured as buy-and-hold-adjusted tracking error improvement.

The scorer quantifies discrete active-delta label errors and computes the improvement metric relative to a no-trade baseline. The patcher module leverages this diagnostic signal—often augmented with time-aligned shareholder commentary—to revise the persona's policy statements and behavioral priors. This iterative reboot—drawing on both observed transition errors and manager-authored rationales—yields a sequence of refined persona checkpoints. Final persona selection is automated via standalone validation on a disjoint set of holdings transitions.

Evaluation Protocol and Dataset

A rigorous experimental protocol evaluates Fund2Persona across both quantitative and qualitative axes, using a universe of 69 funds with strict evidence completeness. Each fund is represented with nine quarterly top-holdings snapshots, regulatory disclosures, mapped price series, and associated market context digests.

Quantitative Metrics

• Copy-trading portfolio reconstruction: For held-out months, each method attempts to infer the next-period portfolio allocations given only prior holdings, market context, and realized returns, within the constraining universe of top-10 prior holdings. The main evaluation metric is active-delta label accuracy and improvement in F1 over deterministic buy-and-hold and other baselines.
• Manager-commentary alignment: For funds with available shareholder commentary, the generated persona writes manager-style memos, which are compared (via blind LLM judges) for semantic and lens alignment with actual manager narratives, focusing on attributions, sector exposures, and positioning rationales.

Baselines

The benchmarks span generic LLMs (no persona conditioning), disclosure-only personas, random out-of-sample personas, and the buy-and-hold no-trade policy benchmark.

Key Results

Numerical Findings

• In portfolio reconstruction, Fund2Persona achieves superior 5-class accuracy (0.339) and macro-F1 (0.259), outperforming all baselines. Notably, the active-delta label recovery for Fund2Persona is $\sim3$ points higher in accuracy than the deterministic baseline, indicating its improved capacity for reconstructing fund-specific decision logic.
• In commentary alignment, Fund2Persona attains a 30.0% rank-1 rate and best overall average rank (2.45), confirming that patching with behavioral and commentary signals produces memos most textually and contextually aligned with the ground-truth manager commentary.

Downstream Diagnostics and Behavioral Claims

• Scenario generation: Conditioning on retrieved, decision-grounded personas (rather than generic prompts) consistently expands the diversity of hypothetical responses to open-ended market queries. Mean intra-set cosine distance increases from 0.313 to 0.381 (91.6% win rate), indicating that retrieved personas robustly broaden the plausible scenario set—crucial for applications in market-forecasting, tail-risk analysis, and agent-based modeling.
• Personalized investor advice: In a simulated advisory environment with 50 realistic investor profiles, matched fund-derived personas achieve the highest turn-level win rates and rubric scores (e.g., 78% win at turn 5, best preference fit and practical usefulness). Mismatched personas and generic LLMs underperform, highlighting the importance of behavioral alignment and profile-to-persona matching for effective, actionable financial guidance.

Theoretical and Practical Implications

Fund2Persona demonstrates that decision-grounded, data-derived personas materially outperform generic or disclosure-only personas in replicating both normative (portfolio allocation) and narrative (manager commentary) elements of financial manager behavior. The agentic refinement loop—grounded in counterfactual replay and real manager rationale—constitutes a scalable protocol for behavioral alignment, distinct from mere style transfer.

On the theory side, this work provides evidence that multi-source formative grounding (holdings delta, market context, ex post commentary) enables LLMs to emulate policy objects that reflect stable, interpretable, and transferable investment logic. This stands in contrast to surface-level persona conditioning, which prior work demonstrates is neither robust nor semantically rich for complex judgment tasks. Fund2Persona thus increases the portability of manager expertise, effectively decomposing investment style, horizon, and risk posture into a format suitable for LLM deployment, simulation, or advisory use.

Practically, this framework enables modular construction of advisor pools reflecting established asset manager philosophies. This both enhances the realism of advisory dialogues and diversifies agent-based scenario modeling, ranging from retail advice to institutional strategy simulation. Moreover, downstream diagnostics suggest these personas can mitigate “heuristic collapse” and overfit to dominant factor exposures—a known failure mode in LLM advisory agents.

Limitations and Future Directions

The main limitations pertain to (i) reliance on specific LLM architectures for both persona generation and evaluation—raising concerns regarding cross-model robustness—and (ii) proxy-based assessment relying on LLM judges for commentary and dialogue evaluation rather than human experts. Future research should expand beyond the current 69-fund strict dataset, diversify LLM architectures, and elicit expert human evaluation to triangulate alignment and practical utility. Additionally, more complex investor constraints (taxes, regulatory restrictions, cross-asset overlays) and multi-agent simulation horizons remain to be explored within this paradigm.

Conclusion

Fund2Persona establishes a rigorous, data-driven pipeline for constructing, refining, and validating financial-advisor personas for LLMs, integrating fund disclosures, holdings transitions, market context, and manager commentary. Empirical results demonstrate that these personas recover fund-specific decision policies and interpretive frameworks more faithfully than generic approaches. The downstream consequence is demonstrably better scenario coverage and investor-specific advisory quality, showing that financial system interfaces can now leverage portable, manager-grounded expertise rather than defaulting to generic, undifferentiated recommendations. The framework offers new directions for both research on persona conditioning and the practice of AI-augmented financial advisory.