Can ChatGPT Beat the Market? Agentic LLM Strategies for Regime-Aware Equity Investment: A Six-Year Multi-Regime Empirical Study of TSMC (2020–2026)

Can ChatGPT Beat the Market? Agentic LLM Strategies for Regime-Aware Equity Investment: A Six-Year Multi-Regime Empirical Study of TSMC (2020–2026)

Title

Can ChatGPT Beat the Market? Agentic LLM Strategies for Regime-Aware Equity Investment: A Six-Year Multi-Regime Empirical Study of TSMC (2020–2026)

Authors

  • Lanz CWJ Chan
    Finamatrix Quant-Lab, Singapore
  • Wing-Keung Wong
    Department of Finance, Quantum AI Research Center, Fintech & Blockchain Research Center, and Big Data Research Center, Asia University, Taiwan
    Department of Medical Research, China Medical University Hospital, Taiwan
    Business, Economic and Public Policy Research Centre, Hong Kong Shue Yan University
    The Economic Growth Centre, Nanyang Technological University, Singapore
    No. 500, Liufeng Road, Wufeng Dist., Taichung City 41354, Taiwan

Abstract

Purpose: This study investigates whether an agentic ChatGPT investment system can generate regime-aware, valuation-disciplined, and capital-constraint-conscious equity strategies that deliver superior risk-adjusted performance relative to passive benchmarks across multiple market regimes.
Design/Methodology/Approach: A five-module agentic pipeline integrates: (i) Hidden Markov Model (HMM) regime classification with geopolitical override; (ii) two-stage discounted cash flow (DCF) valuation anchored to audited FY2025 TSMC financials; (iii) Shannon entropy-minimising structured prompt engineering; (iv) Kelly criterion position sizing with a half-Kelly constraint; and (v) Volume-weighted Golden Ratio Estimator (vGRE) stop-loss management. The pipeline is validated against 55 monthly TSMC (2330.TW) price observations spanning five structurally distinct market regimes from January 2020 to April 2026. Strategy performance is evaluated using return on cost, maximum drawdown, annualised Sharpe ratio, and second-order stochastic dominance (SSD) ordering.
Findings: The Agentic DCA (ADCA) strategy achieves a return on cost of 282.6% versus 274.7% for equal-weight DCA — a 7.9 percentage-point capital efficiency advantage — while reducing maximum drawdown from 18.6% to 16.1%. ADCA establishes second-order stochastic dominance over all passive alternatives for risk-averse investors. During the Iran War Shock regime (March–April 2026), ADCA generates an active return advantage of 997 basis points (approximately 9.97 percentage points) relative to naive buy-and-hold, with a 40.0% reduction in maximum drawdown.
Research Limitations/Implications: Monthly data constrains HMM precision; daily observations (~1,580) would enable formal bootstrap SSD testing. Transaction costs and tax effects are excluded. Single-asset analysis limits generalisability. Results reflect a secular bull market in TSMC driven by the AI infrastructure cycle, which may not persist.
Practical Implications: The five-module framework provides a directly replicable ChatGPT workflow for regime-aware, fundamental-grounded equity analysis across multiple market cycles, with direct applicability to institutional and systematic retail investment.
Originality/Value: This study presents the first empirical validation of a ChatGPT agentic investment pipeline incorporating 44 numbered equations across five empirically verified market regimes spanning six years of real TSMC price data, integrating SSD theory, HMM regime detection, DCF valuation, and Kelly-constrained position sizing within a single coherent decision-theoretic framework.

Keywords

ChatGPT, large language models, agentic AI, dollar-cost averaging, TSMC, semiconductor investment, stochastic dominance, Hidden Markov Model, discounted cash flow, Kelly criterion, geopolitical risk, regime detection

Classification-JEL

C45, C61, G11, G14, G15, G17

Pages

1-29

How to Cite

Chan, L. C., & Wong, W. K. (2026). Can ChatGPT Beat the Market? Agentic LLM Strategies for Regime-Aware Equity Investment: A Six-Year Multi-Regime Empirical Study of TSMC (2020–2026). Annals of Artificial Intelligence, 1, 1-29.