The Anatomy of Market Influence: Evaluating Nonlinear Granger Causality Networks Across The U.S. Equity Market

Abstract

The purpose of this thesis is to investigate whether nonlinear granger causality methods for predictive causal discovery in U.S. equity markets produce directed networks that are more informative and economically viable than comparable linear baselines. The central novelty lies in the empirical design, in which linear Granger, multilayer perceptron (MLP) Neural Granger, and recurrent neural network (RNN) Neural Granger are evaluated on the same fixed 20-asset universe, over the same 2014--2024 sample split, and under the same downstream filtering, lag-refinement, trading-cost, and benchmark framework. Starting from Center for Research in Security Prices (CRSP) data, the thesis applies train-only universe selection, method-specific edge screening, Peter and Clark Momentary Conditional Independence Plus (PCMCI+) filtering, Dynamic Time Warping / k-nearest neighbors (DTW/KNN) lag optimization, rolling Domino Graph diagnostics, and cost-aware pair-trading evaluation with passive and synthetic benchmarks. The findings reveal that nonlinear methods do change the discovered directed network. Relative to the linear Granger benchmark, the nonlinear branches generate more cross-sectoral edges and graphs. In addition, the RNN and MLP edge sets overlap strongly with each other, while overlapping only weakly with the linear edge set. Among the model-based strategies, the RNN branch performs best: it delivers the strongest primary after-cost performance and significantly outperforms both linear Granger and MLP on Sharpe ratio in direct bootstrap comparisons. However, no model-based strategy beats the passive equal-weight benchmark over the full out-of-sample period. Overall, the main conclusion is that nonlinear predictive-causality methods, especially the RNN specification, are more credible as tools for directed market-structure discovery than as standalone engines of passive-beating portfolio performance. The thesis contributes a controlled comparison of linear and nonlinear causal-network methods and a Domino Graph framework capable of showing how asset influence changes over time.