A1270
Title: On the recovery of meaningful structure from financial transfer entropy networks
Authors: Haotian Yang - Bocconi University (Italy) [presenting]
Yifeng Li - University of Trento (Italy)
Abstract: Financial network studies frequently use transfer entropy to identify systemic hubs under high-dimensional constraints, typically at $T/N<5$. A matched data-generating process audit evaluates the node-level estimation reliability of these methods using realistic financial dynamics, including GARCH and common factors. Monte Carlo simulations reveal that at $T/N<5$, ordinary least squares-based estimators yield precision below 17\%, rendering hub detection close to random. LASSO regularization achieves higher precision but collapses to near-empty networks on empirical S\&P 500 data. Furthermore, a statistical power analysis demonstrates that estimation noise alone destroys realistic network-return channels; an embedded 10\% annualized network premium yields an insignificant estimated t-statistic of 0.74 at $T/N=5$, compared to an oracle t-statistic of 5.74. While aggregate connectivity trends partially survive estimation noise, specific node-level topological claims are fundamentally unreliable. The findings establish a binding $T/N$ barrier, indicating that reliable edge-level estimation requires a ratio of at least 8 to 10, necessitating sector-level aggregation or high-frequency data for valid inference.
