Quantitative modelling and learning in financial markets: an agent-based approach

This thesis proposes innovative methodologies for modelling financial markets using agent-based models (ABMs). It addresses key challenges including price formation analysis, optimal hedging strategy, and flash crash analysis. The first contribution is the XGB-Chiarella method, which integrates the Extended Chiarella model with an XGBoost-based calibration technique for realistic financial market simulations. This method, tested across different stock exchanges, effectively replicates market dynamics. The analysis of the price formation process highlights agent-based models' capability to capture universal phenomena in financial markets. The XGB-Chiarella method establishes a new standard for financial market simulations and provides a robust workflow for calibrating ABMs. The second contribution is the creation of the Chiarella-Heston model. This model combines the Extended Chiarella model with the Heston stochastic volatility model. It is calibrated to reproduce empirical stylised facts and is utilised to train a deep hedging agent. Tested against empirical data, the trained agent demonstrates superior performance in optimising hedging strategies compared to traditional models. The Chiarella-Heston model provides financial institutions with a robust tool to enhance risk management strategies. The third contribution is a simulator for simulating high-frequency trading environments and risk scenarios, such as flash crashes. Building on the XGB-Chiarella method, this simulator operates at sub-second intervals with full exchange protocols. It offers a comprehensive tool for analysing the dynamics and conditions leading to flash crashes. The simulator also provides insights into factors that influence the severity of such risk scenarios, including the market maker inventory limits and trading frequency for different types of traders. The simulator also examines mini-flash crash events, enhancing our understanding of high-frequency trading risks. Overall, this thesis demonstrates the effectiveness of an agent-based approach in modelling financial markets, bridging the gap between theory and practice. It provides a foundation for future applications in risk management and policy evaluation.