THE SYSTEMIC FRAGILITY OF DISTILLED GRAPH MODELS IN FINANCIAL MARKETS

The deployment of Graph Neural Networks (GNNs) in financial markets is often constrained by high inference latency, necessitating knowledge distillation from a topological GNN to a structure-agnostic Multi-Layer Perceptron (MLP). This study investigates the systemic consequences of this architectural compression using a TwinMarket-inspired agent-based simulation (Yang et al., 2025) grounded in real-world shocks. We hypothesize that distilling relational intelligence into non-relational MLPs erodes structural awareness, rendering markets fragile to targeted misinformation. Our experiments reveal that while distilled MLPs preserve marginal predictive accuracy, they exhibit a 37% increase in price volatility and suffer from permanent price hysteresis under adversarial belief shocks. Furthermore, we observe that architectural compression facilitates an “unstable meritocracy”, where trading volume centralizes into a narrow elite of structure-blind agents. These findings expose a critical tension: the pursuit of computational efficiency through graph-to-MLP distillation directly compromises the relational inductive biases essential for maintaining systemic market resilience.