Source-linked AI summary
Are LLMs Good Financial User Simulators? Multi-view Investor Logic Alignment (MILA)
Jiajie He, Jiangyuan Hong, Xintong Chen, Dongling Ni, Wenjin Liu
TL;DR
The paper asks whether LLMs reproduce the evolving individual decisions behind financial behavior, rather than merely matching coarse or aggregate actions. It evaluates next-day investor simulation in a longitudinal paper-trading setting and finds that LLMs do not reliably outperform a simple activity-persistence baseline, with fidelity declining at finer decision levels.
Problem
It remains unclear whether an LLM that produces behavior similar to an individual user reproduces that user’s evolving decision process, since longitudinal actions depend on prior behavior, environment state, and earlier consequences.
Source: §1 · p. 1Method
Results
A simple recent-activity persistence rule remains a strong trade-occurrence baseline, while agreement weakens for buy–sell structure and assets, and similar activity predictions can yield different portfolio trajectories.
Source: §6 · p. 11Takeaways & Limitations
Current simulators capture predictive regularities more reliably than stable asset-specific or causal decision mechanisms.
Source: §1 · p. 2, §6 · p. 11Evidence-sensitivity experiments identify what affects simulator outputs but do not establish that the same evidence causally determines real investor behavior.
Source: §3.2 · p. 4
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly used as user simulators, yet it remains unclear whether their predictions faithfully reproduce the evolving decisions of individual users. We investigate this question in a controlled longitudinal paper-trading study with 80 participants, where user interactions, simulated transactions, virtual portfolio states, and point-in-time market information are aligned under a rolling next-day prediction protocol. We evaluate behavioral fidelity hierarchically, from trade occurrence to action structure, asset selection, and downstream portfolio consequences. Across 1,239 aligned user-days, no evaluated LLM reliably outperforms a simple recent-activity persistence baseline for predicting whether a user trades. Fidelity further deteriorates at finer levels: models struggle to recover buy--sell structure and traded assets, and similar activity-level predictions can lead to substantially different portfolio trajectories. Controlled evidence ablations show that recent trading history strongly governs activity prediction, whereas asset selection is substantially more sensitive to the available evidence. An observational analysis further finds that intensified ticker-specific research predicts imminent trading, but diagnostic tests do not support a causal interpretation. These findings suggest that current LLMs capture useful short-term behavioral regularities without yet recovering a stable individual decision mechanism.
1 Introduction
The paper asks whether LLMs reproduce individual investors’ evolving decisions, not merely plausible or aggregate behavior. In a longitudinal paper-trading study, it evaluates fidelity from trade occurrence through portfolio outcomes and finds that coarse prediction does not ensure fidelity to individual decisions.
1 Introduction
- A simulator can match aggregate behavior yet act at the wrong time, select different objects, or respond to different evidence than the individual it represents.Longitudinal decisions depend on prior behavior, environmental state, and consequences of earlier actions.Source: §1 · p. 1
- Prior financial-agent work has emphasized trading performance, recommendation quality, or aggregate market behavior, leaving individual decision-process reproduction comparatively underexplored (Zhang et al., 2024a; YANG et al., 2025; Wang et al., 2026b; Gong et al., 2026).The open question is which aspects of an individual’s decision process are reproduced when an LLM behaves similarly.Source: §1 · p. 1
- The study follows 80 participants in controlled paper trading and uses a rolling next-day protocol restricted to information observable before each prediction cutoff.The trajectories align platform interactions, paper trades, virtual portfolio states, and contemporaneous market information.Source: §1 · pp. 1–2
- Behavioral fidelity is evaluated from whether a user trades to action structure, assets traded, and the resulting portfolio trajectory.The decomposition tests whether coarse agreement also holds at finer decision levels.Source: §1 · p. 2
- Withholding recent trading history greatly degrades activity prediction, while asset predictions shift with nearly any withheld evidence group; ticker-specific research is associated with imminent same-ticker trading, but not causally identified.The findings distinguish evidence affecting simulator predictions from observational associations that do not establish causes of user behavior.Source: §1 · p. 2
- A simple recent-activity persistence rule is competitive with or stronger than evaluated LLMs for trade occurrence, while fidelity weakens for action structure, asset selection, and portfolio trajectories.Similar activity-level predictions can still produce substantially different portfolio trajectories.Source: §1 · p. 2
2 Task Formulation
The task predicts each investor’s structured action on the next U.S. equity trading day using only information available by the prior market close. Its rolling inputs combine investor history, recent behavior, portfolio state, and market information, with inputs reconstructed to prevent future leakage.
2 Task Formulation
- At each market-close cutoff, the simulator is restricted to information observable at or before that time.The target is the next trading day.Source: §2 · pp. 2–3
- The input combines an investor profile, long-term executed trading history, recent seven-day behavior, portfolio state, and market information through the cutoff.Recent behavior includes transactions, views, and news consumption, among other in-platform interactions.Source: §2 · p. 3
- Time-varying inputs are reconstructed independently at each cutoff, and evidence-derived representations are recomputed after ablations to prevent future or cross-group information leakage.The simulator predicts a structured investor action for the next trading day.Source: §2 · p. 3
3 Task Evaluation
The paper frames personalized investor simulation as predicting an individual’s temporally aligned decisions, including abstention, rather than optimizing returns. It evaluates fidelity from trade occurrence through decision structure and evidence sensitivity.
3.1 Hierarchical prediction of investor actions.
- At each cutoff, the simulator predicts a structured next-day action sequence whose components jointly determine the portfolio transition.Source: §3 · p. 3, §3.1 · p. 4
3 Task Evaluation
- The objective is to reproduce an individual investor’s realized decisions, including abstaining from trades, rather than optimize financial returns.Source: §3 · p. 3
3.1 Hierarchical prediction of investor actions.
- The five prediction stages are trade occurrence, day-level action composition, asset selection, position transition, and normalized signed trade scale.Source: §3.1 · p. 3
- Evaluation uses accuracy and macro-F1 for trade occurrence, direction, and position transition, and Hit@5 and NDCG@5 for asset selection.Results are reported stage-wise using ground-truth preceding stages and end-to-end using the simulator’s predictions; asset candidates are point-in-time holdings, previously attended assets, and market-matched distractors.Source: §3.1 · p. 4
3.2 Evaluating investor-specific decision policies.
- Decision-process evaluation tests whether predictions depend on investor-specific evidence through controlled leave-one-group-out interventions and measures changes in probabilities, actions, rankings, or performance.The evidence groups are investor profile, long-term trading history, recent behavior, portfolio state, and market context.Source: §3.2 · p. 4
- Evidence-sensitivity probes identify which evidence affects simulator outputs, but do not establish that the same evidence causally determines real investor behavior.Human evaluation assesses grounding, investor specificity, decision consistency, and historical plausibility; neither it nor interventional faithfulness establishes access to an investor’s unobserved psychological process.Source: §3.2 · pp. 4–5
4 Experiments
The experiments assess financial-simulator fidelity from trade occurrence through action structure, asset selection, and portfolio consequences, then examine which evidence shapes predictions and whether research escalation predicts trading. Across these evaluations, coarse activity agreement does not ensure faithful individual decisions, and the evidence does not establish research as a causal trigger.
4 Experiments
- The study is a pilot aimed at characterizing feasibility and failure modes, not establishing population-level performance.The evaluation distinguishes simulator-input dependence from observational associations in participants' behavior.Source: §4 · p. 5, §4.2 · p. 10
4.1 RQ1: How Faithful Are LLMs as Financial User Simulators?
- The recent-activity persistence baseline achieves 79.82% accuracy and 0.714 macro-F1, and no LLM reliably improves over it.It predicts a trade when the preceding trading day contains an execution.Source: §4.1 · p. 5, Table 1 · p. 6
- The strongest simulator predicts the three-way action side on 129 of 307 trading days (42.02%), only slightly above the 41.37% majority baseline.Buy-and-sell days account for 41.4% of observed trading days, while sell-only recall is 4.23%–8.45% across models.Source: §4.1 · p. 6, Figure 2 · p. 7
- Even the strongest ticker Hit@5 values are around 40–43%, so recognizing a trade or its structure does not ensure recovery of its assets.The 127 joint buy-and-sell days require separate recovery of both ticker sets.Source: §4.1 · p. 6, Table 2 · p. 6
- Candidate pools contain a true bought ticker on only 73.31% of buying days, and ticker errors also persist when useful assets are available.The end-to-end ticker metric reflects both candidate generation and downstream asset selection.Source: §4.1 · pp. 6–7
- User C's activity macro-F1 of 0.627 coexists with real portfolio returns of +11.0% and simulated returns of −8.37%, a 19.37-point gap.Users A and B ended within 0.82 and 1.86 percentage points of their real portfolios, respectively.Source: §4.1 · p. 7, Figure 3 · p. 8
4.2 RQ2: Does the Simulator Recover User Investment Logic?
- Removing recent trade records reduces activity accuracy by 23.49 percentage points, while withholding any evidence block changes more than half of predicted ticker lists.The ablation effects are not additive because evidence blocks overlap.Source: §4.2 · p. 8, Table 3 · p. 9, Figure 4 · p. 9
- Research escalation is associated with a 3.94-point increase in next-day same-ticker trading, but residual imbalance and a positive future-behavior placebo preclude causal interpretation.The study treats research escalation as a temporally proximal behavioral marker rather than an identified causal trigger.Source: §4.2 · p. 9, Figure 5 · p. 10
5 Related Work
Prior work spans general LLM-based human simulation and financial agents, but whether these systems reproduce individual investors’ decisions across scenarios and time remains underexplored.
5.1 LLM as Human Simulator
- Earlier LLM simulators address long-horizon social behavior and web search, while their effectiveness in financial decisions remains largely unexplored.The paper introduces a pilot benchmark combining long-horizon paper-trading behavior with market dynamics.Source: §5.1 · p. 10
5.2 Financial Agent Application
- Financial agents have progressed from recommendation simulators to LLM-based trading agents and evolving-state investor simulators.Examples include StockAgent, TwinMarket, Conv-FinRe, and ShiJianBench.Source: §5.2 · p. 10
- Existing studies have not systematically tested whether simulators reproduce heterogeneous individual behavior across scenarios and extended time horizons.Source: §5.2 · p. 11
6 Conclusion
The paper evaluates whether LLMs reproduce individual financial decisions across behavioral levels and evidence dependencies. It finds that coarse trading agreement does not ensure faithful decisions or stable decision mechanisms.
6 Conclusion
- The study separates fidelity across trade occurrence, action structure, asset selection, and downstream portfolio consequences, while distinguishing predictive performance from evidence dependence.Source: §6 · p. 11
- A simple recent-activity persistence rule remains a strong trade-occurrence baseline, while agreement deteriorates for buy–sell structure and traded assets.Similar activity-level predictions can also produce substantially different portfolio trajectories.Source: §6 · p. 11
- Recent trading history strongly influences activity prediction, but asset selection is substantially more sensitive to changes in available evidence.Intensified ticker-specific research is associated with imminent trading, but the data do not support a causal interpretation.Source: §6 · p. 11
- Simulator evaluation should assess temporal alignment, decision structure, object-level choices, and downstream consequences, while distinguishing predictive dependence from causal evidence.Source: §6 · p. 11
A.1 Baselines
The evaluation uses representative LLMs spanning proprietary and open-weight models, including named GPT and Geimini series.
A.2 Baselines
- The evaluated representative LLMs include both proprietary and open-weight models.Source: §A.1 · p. 12
- The proprietary GPT series includes GPT-5.6 Sol, GPT-5.6 luna, and GPT-5.4o-mini.The passage associates these models with OpenAI (2026), OpenAI (2026), and OpenAI (2024), respectively.Source: §A.1 · p. 12
- The proprietary Geimini series includes Geimini-3-Flash and Geimini-3.5-Flash.The passage associates these models with Google DeepMind (2025) and Google DeepMind (2026), respectively.Source: §A.1 · p. 12
A.2 DataSet
The dataset uses a controlled paper-trading platform rather than participants’ real brokerage data, and records varied simulated investment behavior over three months.
A.2 DataSet
- The study did not access participants’ real brokerage accounts, holdings, credentials, or transaction records.Source: §A.2 · p. 12
- The study recruited 80 volunteers who could trade or remain inactive under real-time market conditions using non-redeemable virtual funds.The paper-trading trajectories were recorded over a three-month observation period.Source: §A.2 · pp. 12–13
- The platform collected more than 230K interactions across market and news browsing, asset viewing, simulated trades, portfolio states, and chatbot use.Source: §A.2 · p. 13
- Records were cleaned and organized chronologically, with earlier user behavior used to construct profiles of preferences, trading patterns, and portfolio characteristics.Viewed financial news was summarized and inserted into the corresponding behavior-sequence positions.Source: §A.2 · p. 13
- The dataset includes active and inactive participants with varied trading styles and financial experience.Source: §A.2 · p. 13
B User Profile Construction
The paper constructs cutoff-safe, versioned profiles from participants’ observable behavioral histories, not inferred psychological traits or ground-truth investment preferences. A staged pipeline normalizes pre-cutoff records, computes auditable features, analyzes them with specialized agents, and validates and synthesizes their outputs.
B User Profile Construction
- Profiles represent observable behavioral history rather than unobserved psychological traits or ground-truth investment preferences.Source: B User Profile Construction · p. 13
- For each prediction cutoff, the pipeline aligns interaction, trade, portfolio, account, and conversation records while excluding target-day and post-cutoff information.Source: B User Profile Construction · p. 13
- A deterministic module computes auditable behavioral statistics directly from records, including event and trade counts, buy and sell amounts, asset concentration, portfolio changes, and holding summaries.Source: B User Profile Construction · p. 13
- Role-specific agents analyze risk, investment logic, goals, short-term trading, and long-term holdings, producing evidence-linked conclusions with uncertainty or missing-evidence indicators.Source: B User Profile Construction · pp. 13–14
- Conflict checking flags inconsistent or unsupported outputs before profile synthesis, while missing observations remain unknown rather than becoming negative behavioral claims.Source: B User Profile Construction · p. 14
- Profiles are versioned by participant and cutoff, with persistent patterns kept in the long-horizon profile and recent conditions represented in separate short-horizon behavior, portfolio, and market-context blocks.Source: B User Profile Construction · p. 14