The List / Agents and Multi-Agent Systems / paper
Strategic Complexity and Behavioral Distortion: Retail Investing Under Large Language Model Augmentation
Abstract
This conceptual article introduces Perceived Cognitive Assistance (PCA)—a novel psychological construct capturing how interactive support from Large Language Models (LLMs) alters investors’ perception of their cognitive capacity to execute complex trading strategies. PCA formalizes a behavioral shift: LLM-empowered retail investors may transition from intuitive heuristics to institutional-grade strategies—sometimes without adequate comprehension. This empowerment–distortion duality forms the theoretical contribution’s core. To empirically validate this model, this article outlines a five-step research agenda including psychological diagnostics, trading behavior analysis, market efficiency tests, and a Behavioral Shift Index (BSI). One agenda component—a dual-agent simulation framework—enables causal benchmarking in post-LLM environments. This simulation includes two contributions: (1) the Virtual Trader, a cognitively degraded benchmark approximating bounded human reasoning, and (2) the Digital Persona, a psychologically emulated agent grounded in behaviorally plausible logic. These components offer methods for isolating LLM assistance’s cognitive uplift and evaluating behavioral implications under controlled conditions. This article contributes by specifying a testable link from established decision frameworks (Theory of Planned Behavior, Technology Acceptance Model, and Risk-as-Feelings) to two estimators: a moderated regression for individual decisions (Equation (1)) and a composite Behavioral Shift Index derived from trading logs (Equation (2)). We state directional, falsifiable predictions for the regression coefficients and for index dynamics, and we outline an identification and robustness plan—versioned, time-locked, and auditable—to be executed in the subsequent empirical phase. The result is a clear operational pathway from theory to measurement and testing, prior to empirical implementation. No empirical results are reported here; the contribution is the operational, falsifiable architecture and its implementation plan, to be executed in a separate preregistered study.
In one sentence
Summary
The paper addresses the behavioral gap in retail investing where Large Language Models (LLMs) lower cognitive barriers to complex strategies, potentially causing an 'illusion of understanding.' It introduces Perceived Cognitive Assistance (PCA) as a psychological construct where investors feel enhanced capability to execute institutional-grade strategies without adequate comprehension. The authors argue this empowerment-distortion duality can lead to behavioral distortion, overconfidence, and emergent market-level risks like algorithmic coherence and volatility amplification among retail cohorts. Methodologically, the paper is conceptual and proposes a five-step research agenda rather than reporting empirical results. It extends the Theory of Planned Behavior (TPB) by integrating the Technology Acceptance Model (TAM) and Risk-as-Feelings Theory. The core methodological contributions are two estimators: Equation (1), a moderated regression linking LLM engagement, PCA, and behavioral outcomes, and Equation (2), a composite Behavioral Shift Index (BSI) derived from trading logs to measure shifts in strategy complexity. The paper also outlines a dual-agent simulation framework using a 'Virtual Trader' (bounded cognition) and a 'Digital Persona' (behaviorally plausible logic) for causal benchmarking in future studies. The findings are theoretical predictions rather than empirical results. The authors posit that PCA moderates the relationship between LLM use and complex strategy adoption, with positive interaction effects expected. Limitations include the lack of current empirical data, the reliance on future preregistered studies, and the challenge of isolating LLM effects from other market factors. The paper serves as a foundational framework for measuring cognitive shifts in AI-augmented retail trading, highlighting the need for investor protection policies and platform design changes to mitigate behavioral risks.
Research problem
The paper investigates how LLM augmentation affects retail investors' strategy selection and risk posture, specifically focusing on the behavioral shift from intuitive, low-complexity strategies to complex, high-risk institutional-grade tactics. It addresses the lack of theoretical frameworks explaining the psychological mechanisms behind this shift, particularly the 'illusion of understanding' where perceived competence rises without actual skill acquisition. The core problem is distinguishing between genuine empowerment and behavioral distortion caused by LLM-mediated cognitive scaffolding.
Core contributions
- Introduction of Perceived Cognitive Assistance (PCA) as a novel psychological construct extending Perceived Behavioral Control in the Theory of Planned Behavior.
- Development of a moderated regression model (Equation 1) linking LLM engagement, PCA, and behavioral outcomes to test causal mechanisms.
- Proposal of a composite Behavioral Shift Index (BSI) derived from trading logs to quantify changes in strategy complexity and risk exposure over time.
- Design of a dual-agent simulation framework featuring a 'Virtual Trader' and 'Digital Persona' for causal benchmarking of LLM impacts.
- Integration of Risk-as-Feelings Theory and Technology Acceptance Model (TAM) to explain affective divergence and technology adoption heterogeneity.
Method
- Extends Theory of Planned Behavior (TPB) by adding Perceived Cognitive Assistance (PCA) as a moderating construct.
- Integrates Technology Acceptance Model (TAM) constructs (Perceived Usefulness, Perceived Ease of Use) as upstream determinants of LLM engagement.
- Applies Risk-as-Feelings Theory to model affective divergence between cognitive intention and behavioral execution.
- Proposes Equation (1): Bi(t) = β0 + β1LLM + β2PCA + β3(LLM×PCA) + β4Attitude + β5SubjectiveNorms + Controls + ε.
- Proposes Equation (2): A composite Behavioral Shift Index (BSI) aggregating period-over-period changes in trading metrics.
- Designs a dual-agent simulation with a Virtual Trader (bounded cognition) and Digital Persona (behaviorally plausible logic) for causal benchmarking.
- Plans for preregistered, version-controlled, and time-locked empirical implementation to ensure auditability.
Results & findings
- PCA is defined as the subjective belief in enhanced cognitive capability to execute complex tasks via LLM support, distinct from actual performance.
- The paper predicts a positive interaction effect between LLM engagement and PCA on the adoption of complex strategies.
- Behavioral Shift Index (BSI) is proposed to aggregate changes in multi-leg share, trading frequency, concentration, and volatility exposure.
- Risk-as-Feelings Theory suggests that emotionally charged LLM narratives may cause intention-behavior divergence.
- Theoretical predictions include overconfidence and confirmation bias as key risks associated with high PCA.
- The study outlines a plan for versioned, time-locked, and auditable empirical implementation to ensure robustness.
Data & experiments
- Conceptual paper proposing a five-step empirical research agenda; no current empirical results are reported.
- Proposed data sources include investor-level trading logs, psychological diagnostics, and market efficiency metrics.
- Experimental design involves a preregistered, version-controlled study to be executed in a subsequent phase.
- Simulation framework uses synthetic baselines: Virtual Trader (bounded cognition) and Digital Persona (behaviorally plausible logic).
- Evaluation metrics include regression coefficients for behavioral shifts and index dynamics of the Behavioral Shift Index (BSI).
Limitations & caveats
- No empirical results are reported; the paper is purely conceptual and proposes a future research agenda.
- The validity of PCA and the BSI depends on future empirical validation, which has not yet been conducted.
- The dual-agent simulation is a proposed method, not an implemented or tested system.
- Theoretical predictions are directional and falsifiable but lack quantitative support in this publication.
- The paper acknowledges the challenge of isolating LLM effects from other market factors in real-world data.
- Reliance on self-reported psychological diagnostics may introduce measurement bias in future empirical studies.
Curator notes
Discovery status: institutional_trading_round3_promoted. Source links: ABIDES: Towards High-Fidelity Multi-Agent Market Simulation