金融市场中的三元压力指数
The Triadic Stress Index in Financial Markets
AI总结：
该研究将源自土壤微生物组共现网络的三元压力指数应用于金融关联网络，在多市场样本测试中，其逐节点分解能力、警报清晰度及F1得分均优于多数基准指标，为金融危机监测提供了更优工具。
AI中文摘要：
三元压力指数（Triadic Stress Index, TSI）是一个网络指标，其四个因子最初在土壤微生物组共现网络中被观测到，且未做任何改动便被应用于金融资产的关联网络。我们在2006-2026年涵盖五个市场（包括含银行危机和AI领域的股票、加密货币、大宗商品、外汇及主权债务）的样本上对其进行测试，采用固定警报预算、样本外测试，结合块自举区间及多重检验的Holm校正，针对三种独立的危机事件定义展开评估。基准指标包括MSCI及央行使用的行业标准——吸收比率（Absorption Ratio）、现有最尖锐的谱测度——有效秩（effective rank）和Vendi分数、Ollivier-Ricci曲率，以及符号关联网络的全局和局部平衡指数。三项对比结果支持该指数：它具备基于diag(A^3)的逐节点分解，可明确指出承载集中度的资产且无需选择参数，其得分在0.97-0.99之间，而唯一已发表的逐节点替代指标得分仅为0.33-0.84，谱归因则需先选择要读取的成分数量，且在标准但错误的选择下会失效；其警报是所有测试指标中最清晰的，仅4.0%的警报无对应事件，而有效秩的这一比例为14.7%，吸收比率约为59%；在样本外F1得分上，它比吸收比率高出0.273，p值小于0.0005。其余对比结果为平局：在最严格的标签下，其与吸收比率的优势缩小，而在所有方案及两个样本中，它与有效秩和Vendi分数均为平局。在真实矩阵上，远更简单的节点度可复现该归因。领先-滞后分析显示峰值互相关在零滞后处，因此这是一个同步状态指数，而非预测指标。
英文摘要：
The Triadic Stress Index (TSI) takes a network index whose four factors were first observed in soil microbiome co-occurrence networks and applies it, without alteration, to the correlation network of financial assets. We test it on five markets spanning 2006-2026 (equities including banking crises and the AI sector, cryptocurrencies, commodities, foreign exchange and sovereign debt), against three independent definitions of a crisis episode, at a fixed alarm budget, out of sample, with block-bootstrap intervals and a Holm correction across the family of tests. The benchmarks are the Absorption Ratio, the industry standard used by MSCI and central banks; the effective rank and the Vendi score, the sharpest spectral measures available; Ollivier-Ricci curvature; and the global and local balance indices of signed correlation networks. Three comparisons favour the index. It carries a per-node decomposition, diag(A^3), naming which asset is carrying the concentration with no parameter to select, and scores 0.97-0.99 against 0.33-0.84 for the only published per-node alternative, whereas spectral attribution must first choose how many components to read and collapses under a standard but wrong choice. Its alarms are the cleanest of anything tested, 4.0% of them with no matching episode against 14.7% for the effective rank and roughly 59% for the Absorption Ratio. And it beats the Absorption Ratio on detection by 0.273 in F1 out of sample, p<0.0005. The remaining comparisons are ties. Against the effective rank and the Vendi score the index ties in every scheme and both samples, and the margin over the Absorption Ratio narrows under the strictest labelling. On real matrices the far simpler node degree reproduces the attribution. A lead-lag analysis puts the peak cross-correlation at zero lag: this is a coincident state index, not a forecast.
