AI-Driven Dynamic Asset Pricing with Regime Switching and Nonlinear Risk Premia

Traditional asset pricing models typically assume that systematic risk exposures and risk premia remain constant over time. This paper extends the traditional framework by developing a machine learning approach to regime-dependent dynamic asset pricing that integrates Gaussian Hidden Markov Models (HMMs) with conditional asset pricing theory to identify latent economic states and estimate state-dependent risk compensation, thereby providing a foundation for nonlinear AI-based asset pricing models.

Using monthly macro-financial variables—including equity market returns, realized volatility, credit spreads, yield-curve measures, and short-term interest rates—we identify latent market regimes over the period 1991–2025. The estimated HMM classifies market conditions into three economically interpretable states: Expansion, Normal, and Stress. These regime labels are then merged with monthly returns of broad-market, style, and sector exchange-traded funds (ETFs) to examine whether investment performance and systematic risk vary across latent economic conditions.

Within each regime, we estimate annualized returns, Sharpe ratios, maximum drawdowns, conditional CAPM betas, and the cross-sectional market price of risk. The empirical results reveal substantial regime dependence across all performance measures. Growth-oriented and technology-focused ETFs exhibit stronger risk-adjusted performance during Expansion and Normal regimes, whereas defensive sectors experience smaller drawdowns during Stress periods. Conditional CAPM betas differ across regimes, suggesting that systematic market risk exposures are state dependent rather than constant through time. Furthermore, the estimated market price of risk is positive and statistically significant during Expansion and Normal regimes, while the corresponding estimate is negative during Stress periods, providing evidence that investors price systematic risk differently across economic environments.

These findings indicate that latent market regimes contain economically meaningful information beyond that captured by traditional linear asset pricing models and support the development of regime-conditioned multifactor models and AI-driven nonlinear stochastic discount factor frameworks for dynamic asset pricing.