GeomHerd: A Forward-looking Herding Quantification via Ricci Flow Geometry on Agent Interactive Simulations

Battle Count: 56

Highly relevant to quantitative trading through forward-looking herding detection that precedes price-correlation signals by 40+ steps. The curvature-conditioned forecasting head reduces cascade-window log-return MAE over price-only baselines, directly applicable to regime-aware trading strategies. The contagion-bridge detector (beta-) provides early warning of systemic cascades. However, the agent interaction graph is not directly observable in real markets without full action-stream access, limiting immediate deployment. The 13F retrospective template provides a path to practical implementation under disclosure constraints.

High complexity due to multiple interacting components: (1) LLM-driven multi-agent simulation with persona-conditioned agents, (2) dynamic graph construction with windowed agreement and sparsification, (3) exact Ollivier-Ricci curvature computation via linear-programming Wasserstein distance (POT library), (4) discrete Ricci flow simulation for neckpinch time, (5) CUSUM + Kendall-tau detection with calibration, (6) FSQ codebook for effective vocabulary, (7) Kronos-style discrete tokeniser + Transformer forecasting head with AdaLN-Zero conditioning, (8) mean-field bridge derivation. Requires expertise in differential geometry, optimal transport, time-series detection, LLM orchestration, and financial modeling.

Code and configurations are available at an anonymous link. The paper provides detailed hyperparameters (window T_w=100, sparsification threshold w_0=0.5, snapshot frequency Delta_t=10, laziness alpha=0.5, curvature thresholds kappa+=0.1, kappa-=0.1, CUSUM parameters). Full calibration sweep (Table 8) and ablation studies (Table 4) are provided. Seeds and output hashes are archived. However, LLM prompt corpora are intentionally withheld to mitigate adversarial replication, and the CWS substrate is heavily stylized.