気候移行リスクと世界金融脆弱性：マルチソース説明可能機械学習・パネル計量経済・ストレスシナリオ分析

Climate Transition Risk and Global Financial Vulnerability: A Multi-Source Explainable Machine Learning, Panel Econometric, and Stress-Scenario Analysis (原題)

Wisaeng K, Zheng H

Research Squareプレプリント2026-11-10#気候リスク経営インパクト: 資金調達対象セクター: finance

DOI: 10.20944/preprints202611.1038.v1

原典: https://doi.org/10.20944/preprints202611.1038.v1

🤖 gxceed AI 要約

日本語

世界銀行・OWIDデータを用い、111カ国2001–2022年のパネルで気候移行リスク指数（CTRI）と金融脆弱性指数（GFVI）の関係を検証。固定効果推定ではCTRI係数はβ=0.228（p=0.094）と10%水準で限界的、信用脆弱性指標には有意（β=0.261）。時系列予測ではElastic Netが最良で非線形MLの優位性は不支持、SHAPでは遅行銀行資本・インフレ・GDP成長が支配的でCTRI寄与は約4.2%。

English

Using World Bank and OWID data for 111 countries (2001–2022), the study tests whether a Climate Transition Risk Index predicts a Global Financial Vulnerability Index. Fixed-effects estimates show only marginal significance (β=0.228, p=0.094), though transition risk is significantly linked to a credit-vulnerability measure. Chronological prediction favors Elastic Net over nonlinear ML, and SHAP shows lagged CTRI contributes only ~4.2% of predictive importance.

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

気候移行リスクをマクロ金融安定性と結びつける実証枠組みは、日本銀行や金融庁が進める気候ストレステスト・移行リスク計量化の議論に直接示唆を与える。SSBJ開示や有報でのシナリオ分析高度化を検討する日本企業・投資家にとって、指標設計の感応度と限界を理解する材料となる。

In the global GX context

This paper speaks to the global push—NGFS scenarios, TCFD/ISSB transition-risk disclosure, and central-bank climate stress testing—to quantify how transition exposure transmits to financial stability. Its finding that results hinge on measurement choices and that nonlinear ML offers no clear edge is a useful caution for disclosure-infrastructure and scenario-analysis design under CSRD and ISSB regimes.

👥 読者別の含意

🔬研究者:移行リスク指標の構築とML予測手法の比較、SHAPによる要因分解の実証例として参照できる。

🏢実務担当者:移行リスクが自社の資金調達・資本コストに及ぼす経路を理解し、シナリオ分析の前提設計に活用できる。

🏛政策担当者:気候ストレステストや移行リスク計量化において、指標選択と非線形モデル依存への注意点を示す。

📄 Abstract（原文）

Climate transition risk is increasingly relevant to macro-financial stability, but its empirical effects may depend on the way transition exposure and financial vulnerability are measured. This study integrates actual country-year data from the World Bank World Development Indicators and Our World in Data CO2 database to examine climate transition risk using panel econometrics, chronological out-of-sample prediction, and explainable machine learning. After harmonization, winsorization, and complete-case screening, the main empirical panel contains 1,292 observations for 111 countries over 2001–2022. The Climate Transition Risk Index (CTRI) combines carbon intensity, energy intensity, fossil-fuel dependence, and inverse renewable-energy use, whereas the Global Financial Vulnerability Index (GFVI) combines inflation, weak GDP growth, and low bank-capital buffers. Country- and year-fixed-effects estimates show a positive CTRI coefficient (β=0.228, p=0.094), providing only marginal evidence at the 10% level rather than conventional 5% significance. Transition risk is significantly associated with a separate credit-vulnerability measure (β=0.261, p=0.015), but the macro-financial fragility interaction, private-credit interaction, quadratic CTRI term, and income-group heterogeneity are not statistically supported. In chronological prediction for 2019–2021, Elastic Net records the lowest test RMSE (0.979), narrowly outperforming the ensemble and other linear and nonlinear learners; therefore, the available data do not support the proposed superiority of nonlinear machine learning. SHAP analysis of XGBoost indicates that lagged bank capital, inflation, and GDP growth dominate predictive importance, while lagged CTRI contributes about 4.2%. Because the supplied data included no NGFS scenario file, we retain forward-looking scenario hypotheses as a future extension rather than report them with simulated values. The revised evidence highlights the importance of separating empirical findings from provisional scenario assumptions and shows that climate-transition effects are more nuanced than the original simulated manuscript structure implied.

🔗 Provenance — このレコードを発見したソース

Research Square https://doi.org/10.20944/preprints202611.1038.v1first seen 2026-11-13 04:30:50 · last seen 2026-11-14 04:23:27

🔔 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。