Algorithmic portfolios on Pakistan's stock exchange outperform human-managed accounts on risk-adjusted returns and downside protection, but hybrid investors erode those gains by overriding AI after early wins—fueling overconfidence and larger losses during market stress.

Main Finding

AI-driven portfolios on the Pakistan Stock Exchange (PSX) display lower turnover, better diversification and lower downside risk than human-only portfolios, which trade excessively and earn lower risk‑adjusted returns. However, hybrid human–algorithm investors often override algorithmic advice after early successes, producing a learning-based form of overconfidence (illusion of control) that erodes the efficiency and risk benefits of AI—particularly during financial stress. In short, AI does not eliminate behavioral biases; it reshapes how overconfidence manifests in modern, mixed decision environments.

Key Points

• Overconfidence is operationalized with three empirical proxies: turnover, portfolio concentration (holding bias), and the relative deviation of human trades from algorithmic recommendations.
• Human-only investors: higher turnover, more concentrated portfolios, worse risk‑adjusted performance.
• Algorithm-only investors: better diversification, lower downside risk, superior outcomes in well-structured, data-rich contexts.
• Hybrid investors: tend to follow algorithms initially but increasingly exercise discretionary overrides after early gains, consistent with (i) learning-based confidence increases and (ii) illusion of control; these overrides are associated with higher volatility and deeper drawdowns in stressed markets.
• Human–algorithm interaction produces two opposing behavioral patterns: algorithm appreciation (over-reliance/automation bias) and algorithm aversion (distrust/overriding). Overconfidence helps explain why both can coexist.
• Systemic concerns: widespread use of similar algorithms can create common-mode errors; discretionary interference by overconfident humans can amplify volatility and systemic fragility.
• The paper extends classical overconfidence theory into AI-mediated markets and highlights that behavioral biases persist, but their economic effects differ when algorithms are involved.

Data & Methods

• Data period and market: Pakistan Stock Exchange (PSX), 2020–2025.
• Primary data sources:
  • Anonymized transaction-level retail investor data from a large online brokerage (panel of active domestic individual accounts meeting minimum activity thresholds).
  • Algorithmic recommendations generated by robo-advisors / rule-based or machine‑learning portfolio models (mean–variance and ML-based systems described).
  • Market returns, volatility measures and benchmark indices from established financial databases.
• Decision regimes compared:
• Human-only (pure discretionary trading).
• Algorithm-only (decisions made exclusively by AI/rule systems).
• Hybrid (humans receive algorithmic recommendations and may accept, modify or reject them).
• Empirical strategy:
  • Constructed investor-level panel measures for turnover, concentration, deviation from algorithmic advice, downside risk and risk‑adjusted returns.
  • Regression analyses (panel regressions controlling for market conditions, investor sophistication and risk preferences) to identify associations between overconfidence proxies, regime type, and performance/volatility outcomes.
  • Focused tests on behavior during periods of financial stress to evaluate drawdown and volatility amplification.
• Robustness / identification: multiple overconfidence proxies used; comparison across the three regimes to isolate interaction effects. (Paper reports regressions supporting the main findings; detailed coefficient estimates and exact model specifications are provided in the full text.)

Implications for AI Economics

• Behavioral persistence: AI does not automatically remove investor biases. Economists and practitioners must treat algorithmic adoption as altering, not eliminating, behavioral dynamics.
• Market efficiency vs. fragility: While algorithms can improve diversification and reduce individual downside risk, human discretionary overrides driven by overconfidence can increase market volatility and mispricing—especially during stress—weakening aggregate efficiency gains.
• Distributional effects: If retail investors react to AI with misuse (overriding or blind trust), return disparities between institutional and retail investors may widen; platforms enabling retail use of algorithmic tools should monitor outcomes.
• Systemic risk & common algorithms: Widespread reliance on similar AI models raises the risk of correlated failures; human overconfidence in overriding or trusting these models complicates regulation and stress testing.
• Policy and design recommendations:
  • Improve transparency and auditability of algorithmic recommendations; require disclosure of model limitations and historical performance.
  • Design human–AI interfaces and decision protocols that limit harmful discretionary overrides (e.g., require documented rationale for overrides, cooling-off periods, or algorithmic override caps during volatile periods).
  • Mandate scenario-based stress tests for widely used investment algorithms and monitor common exposures across institutions.
  • Educate retail investors about algorithmic strengths and limits to reduce both automation bias and counterproductive overconfidence.
• Research directions: cross-country comparisons, longer-horizon studies, experiments to causally identify mechanisms behind hybrid overrides, and study of institutional versus retail differences in human–AI interactions.