Information Propagation Across Investor Types: Transfer Entropy Networks in the Korean Equity Market

Battle Count: 72

The paper is highly relevant to quantitative trading as a negative result: it demonstrates that TE network centrality in investor-type flow networks does NOT provide exploitable alpha at the daily frequency for large-cap Korean equities. This constrains strategy design by showing that (1) combining multiple investor-type signals yields redundant rather than synergistic information, (2) network centrality adds negligible cross-sectional predictive power, and (3) the Kelly-optimal growth rate equals the risk-free rate. While the finding is negative, it is practically valuable for avoiding wasted research effort on network-based investor-flow strategies at daily frequency. The methodology (TE networks, interaction information, Kelly bounds) is directly applicable to other markets and frequencies.

Implementation requires: (1) symbolic TE computation with quantile discretization across 9,900 directed pairs per investor type, (2) 200 block-permutation surrogate tests per pair with FDR correction, (3) network centrality measures (out-degree, in-degree, betweenness, closeness, PageRank), (4) interaction information and conditional TE estimation using KSG estimator, (5) Kelly criterion and Fano inequality calculations, (6) Fama-MacBeth cross-sectional regressions over 1,231 days, and (7) multiple robustness checks. The computational burden is moderate (100 stocks, 1,231 days) but the statistical rigor requires careful implementation of permutation tests and multiple comparison corrections. No code is provided.

The paper provides detailed methodology including symbolic TE estimation parameters (5 quantile bins, k=1, l=1), block-permutation surrogate procedure (200 surrogates, block size 20), FDR correction (Benjamini-Hochberg, alpha=0.05), and Fama-MacBeth regression specification. However, no code repository is mentioned, and the specific signal construction (S_MC, S_TV) references a prior working paper (Kang 2025) that may not be publicly available. Data from KRX is publicly accessible but requires specific processing.