Agentic Finance: An Adaptive Inference Framework for Bounded-Rational Investing Agents

Area of Science:

• Computational Neuroscience
• Financial Mathematics
• Economics

Background:

• Financial markets are inherently non-stationary, posing challenges for traditional portfolio management.
• Active Inference offers a framework for understanding goal-directed behavior under uncertainty.
• Existing models often rely on fixed objectives, failing to adapt to changing market dynamics.

Purpose of the Study:

• To introduce Adaptive Inference, a framework extending Active Inference to non-stationary financial environments.
• To model investment decisions as coupled dynamics of belief updating, preference encoding, and action selection.
• To investigate the performance of adaptive versus passive strategies in managing portfolio risk.

Main Methods:

• Developed a portfolio management framework integrating inference, control, and execution under endogenous uncertainty.
• Modeled portfolio behavior via expected free energy (EFE) minimization.
• Conducted train-test evaluation on the ARKK Innovation ETF (2015-2025) and cross-asset validation on the S&P 500 ETF (SPY).

Main Results:

• Identified a 'Passivity Paradox' where frozen belief transfer outperformed naive adaptive learning.
• A Professional Agent achieved a Sharpe ratio of 0.39, while its adaptive counterpart yielded -0.28.
• All agents reduced realized volatility compared to a Buy-and-Hold strategy, with inference-guided risk shaping yielding a positive Entropic Sharpe Ratio (ESR).

Conclusions:

• Adaptive Inference provides a robust approach to endogenous risk management in ambiguous financial regimes.
• The framework mitigates overtrading and distributional shifts by implementing uncertainty-aware passive investing.
• Inference-guided risk shaping quantifies the economic value of information under thermodynamic constraints.