線形仮定を超えて：企業における気候起因の金融伝染のニューラルネットワーク追跡
Beyond linear assumptions: neural network tracing of climate-driven financial contagion in firms (原題)
Tong Niu, Si-Yu Zhu, Lu-Lu Qin
🤖 gxceed AI 要約
日本語
物理的・移行気候ショックから企業デフォルト確率への伝達経路を、解釈可能なANN（MLP＋SHAP）で定量化した研究。北米・欧州・アジア太平洋の上場1,232社（2014–2023）で四半期先PD予測AUCは0.84と線形ロジット（0.72）を上回る。洪水・レバレッジ・干ばつが主要ドライバーで、洪水影響は0.5m超で非線形に急増し高レバレッジ企業で6.8倍。炭素政策ショックは即時、洪水は2四半期後に最大となる。
English
An interpretable ANN (MLP + SHAP) quantifies transmission pathways from physical and transition climate shocks to corporate default risk. For 1,232 listed firms across North America, Europe, and Asia-Pacific (2014–2023), one-quarter-ahead PD prediction reaches AUC 0.84 versus 0.72 for linear logit. Flood depth, leverage, and drought are top drivers; flood impacts are strongly non-linear above 0.5 m and 6.8× steeper for high-leverage firms. Carbon-policy shocks hit immediately, floods peak after two quarters.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
SSBJ・TCFD開示が求める気候リスクの定量化・シナリオ分析に対し、非線形な金融伝達を扱う手法を提供する。日本の金融機関・事業会社がストレステストや移行計画の高度化に活用しうる。
In the global GX context
Speaks directly to TCFD/ISSB/CSRD-driven climate risk quantification and stress testing, offering a non-linear, explainable alternative to linear credit models. Relevant to banks and regulators building transition-risk and physical-risk scenario analysis.
👥 読者別の含意
🔬研究者:気候金融リスクの非線形伝達を解釈可能MLで扱う方法論的枠組みを提供する。
🏢実務担当者:与信・リスク管理部門が気候シナリオをPDに織り込む際の非線形モデル設計に参考となる。
🏛政策担当者:気候ストレステストの手法高度化や開示要件の定量化を検討する際の示唆を与える。
📄 Abstract（原文）
While climate shocks are often a linear process in the physical world, they are more complex in the corporate financial world, as the non-linearity of transmission channels and firm heterogeneity are poorly captured by existing linear models. We create an interpretable Artificial Neural Network framework that is composed of a multi-layer perceptron and SHAP to quantify the transmission pathways from physical shocks and transition shocks to corporate default risk. Our ANN has an out-of-sample AUC of 0.84 for one-quarter-ahead prediction of PD for 1,232 listed firms from North America, Europe, and Asia-Pacific (2014–2023) compared to linear logit (0.72). The top three drivers, according to SHAP analysis, are flood depth (21.3 bps), leverage (14.7 bps), and drought (11.2 bps). Flood impacts are strongly non-linear, increasing rapidly above 0.5 m, and the slope for high-leverage firms is 6.8×. Technology’s industry vulnerability score is between 18.0 bps and 3.0 bps. Direct asset damage is responsible for 58% of flood impacts and is expected to be responsible for 60% of carbon-policy shocks. Spatial falsification (placebo) tests indicate the results are not attributable to model artifacts or spatially correlated confounders (mean SHAP 0.053 bps, p = 0.17). Temporal impulse response results show that the effects of flooding are maximum at 14% after 2 quarters, the effects of carbon are immediate (9%), and the effects of drought are maximum at 12% after 3 quarters. The ANN has an R 2 of 0.914 on the test set and well-behaved residuals (mean of −1.51 bps). Our framework offers a granular, non-linear and explainable approach to quantifying climate-to-financial risk transmission and has direct applications to stress testing and risk management.
🔗 Provenance — このレコードを発見したソース
- semanticscholar https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2026.1909195/pdffirst seen 2026-10-12 05:20:31 · last seen 2026-10-14 05:51:27
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。