Calibrating emergent phenomena in stock markets with agent based models

Area of Science:

• Computational economics
• Financial modeling
• Artificial intelligence in finance

Background:

• Agent-based models (ABMs) offer a frontier for economics, capturing complex dynamics often missed by standard models.
• Widespread adoption of ABMs is limited by a lack of reliable calibration methods.
• This study proposes a novel calibration approach focusing on abnormal trading performance.

Purpose of the Study:

• To develop and test a new calibration method for agent-based models (ABMs) using real financial data.
• To assess the investment performance of ABMs compared to a buy-and-hold strategy.
• To explore the potential of ABMs for creating market crisis warning systems.

Main Methods:

• Equivalence proof between standard minority/majority agents and optimal decision trees.
• Exhaustive testing of single-agent models for anomalous investment performance on NASDAQ data (20 years).
• Principal component analysis for model convergence and development of a two-agent model.

Main Results:

• Significant predictive power and anomalous Sharpe ratio/directional accuracy identified in single-agent models.
• Enhanced performance observed during market bubbles (dotcom) and crises (2008).
• A novel two-agent model demonstrated superior investment performance, particularly during volatile periods.

Conclusions:

• Optimal decision trees provide an efficient representation for calibrating ABMs.
• Calibrated ABMs exhibit strong predictive capabilities and can achieve abnormal trading performance.
• This approach facilitates the development of advanced warning systems for market crises through emergent collective intelligence.