Causal Discovery Algorithms for Improved Decision Making

Abstract

This thesis investigates the use of causal discovery algorithms for predicting financial time series, focusing on three main models: the Equivalence-based Search Strategy, the Peter-Clark Strategy, and Structural Modeling. These models are evaluated using synthetic data generated from Geometric Brownian Motion (GBM) and real stock market data concerning normalized closing prices. The synthetic data, designed to mirror the volatility and autocorrelation observed in global financial markets, were validated for realism using Kolmogorov-Smirnov tests, displaying a mean of 0.045 and a standard deviation of 0.23. Additionally, real stock market data was standardized for consistent scale comparison. The effectiveness of these models was assessed through three quantitative metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Structural Intervention Distance (SID). Experimental results indicate that while the models demonstrated moderate predictive accuracy, with the Peter-Clark Algorithm showing the lowest MAE and RMSE for both synthetic and real data, they effectively reconstructed causal relationships in the synthetic setting. Furthermore, the models displayed robust noise tolerance, particularly in real data scenarios, suggesting potential for practical applications in financial analysis.

Date of Award	30 Dec 2024
Original language	American English
Supervisor	Mohamed Zemerly (Supervisor)