← 論文一覧に戻る

強化学習に基づくESG格付け補正と偽善罰則メカニズムのモデル設計分析

Analysis of Model Design for ESG Rating Correction and Hypocrisy Punishment Mechanism Based on Reinforcement Learning (原題)

*

Zenodo (CERN European Organization for Nuclear Research)データセット2026-10-20#AI×ESGOrigin: CN経営インパクト: 資金調達対象セクター: finance
DOI: 10.5281/zenodo.23513365
原典: https://doi.org/10.5281/zenodo.23513365

🤖 gxceed AI 要約

日本語

中国A株862社・2018〜2023年の四半期パネルを用い、強化学習(PPO)でESG格付けの補正と企業の「偽善」(開示と実態の乖離)への罰則を同時に設計する枠組みを提案。21次元状態・複合報酬・actor-criticネットワークを構築し、静的・教師あり・グラフ・最適化・RLの各ベースラインと比較。再現パッケージと全表・図の生成コードを公開している。

English

Proposes a reinforcement-learning (PPO) framework that jointly corrects ESG ratings and penalizes corporate hypocrisy, using a 15,786 firm-quarter panel of 862 Chinese A-share firms (2018-2023) rated by Wind, Bloomberg and Refinitiv. A 21-dim state, compound reward and actor-critic network are benchmarked against static, supervised, graph, optimization and RL baselines, with a full reproducibility package.

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・有報開示とESG格付けの乖離が国内でも論点となる中、格付け補正と偽善検知をAIで自動化する設計は、日本企業の開示信頼性評価や格付会社・投資家の検証プロセスに示唆を与える。

In the global GX context

Speaks directly to global debates on ESG rating divergence and greenwashing under CSRD/ISSB, offering an ML mechanism to penalize disclosure-reality gaps and improve rating integrity for regulators and data providers.

👥 読者別の含意

🔬研究者:RLとESG評価を接合する再現可能なベンチマーク設計として、格付け補正・偽善検知の手法研究に活用できる。

🏢実務担当者:自社の開示と実態の乖離が格付け補正や罰則の対象になり得る点を、開示ガバナンス設計に反映できる。

🏛政策担当者:格付けのばらつきとグリーンウォッシュ抑制に向けたAI検証・罰則メカニズムの政策設計に示唆を与える。

📄 Abstract(原文)

# Analysis of Model Design for ESG Rating Correction and Hypocrisy Punishment Mechanism Based on Reinforcement Learning Reproducibility package for the manuscript of the same title. The repository holds two archives: * `esg-rating-correction-rl-code.zip` — the Python implementation of every method of Section 3, every baseline of Section 4.1 and the orchestration script that writes every table and figure of Section 4. * `esg-rating-correction-rl-data.zip` — the self-built firm-quarter panel documented in Section 4.1, together with the external hypocrisy evidence and the annotation materials behind the text features. Download both archives, extract them side by side and merge the `data/` folder of the data archive into the code archive before running the scripts. ## Study at a glance | Item | Value | |---|---| | Market | China A-share listed companies | | Period | 2018Q1 to 2023Q4, 24 quarters | | Firms | 862 | | Firm-quarter observations | 15,786 (unbalanced) | | Rating agencies | Wind, Bloomberg, Refinitiv | | Shenwan level-1 industries | 28 | | Training / validation / test | 10,528 / 2,634 / 2,624 | | External hypocrisy labels | 1,174 firm-quarters (800 / 190 / 184) | | Marking threshold θ | 0.08 | | State dimension | 21 (K=3, M=12, P=4, action history 2) | | Actor-critic parameters | 64,067 | ## Software stack The manuscript (Section 3.4) specifies the following stack. The code works on Linux and macOS; a 64-bit Python interpreter is required. ``` Python 3.10 PyTorch 2.1 (CPU or CUDA 11.8) Stable-Baselines3 2.2.1 Gymnasium 0.29 NumPy 1.26, pandas 2.1, scikit-learn 1.3, SciPy 1.11 statsmodels 0.14, xgboost 2.0 matplotlib 3.8, PyYAML 6.0 ``` Install the dependencies with ``` pip install -r requirements.txt ``` ## Repository layout ``` esg-rating-correction-rl/ ├── README.md top-level description (this file) ├── requirements.txt pinned dependencies ├── method_fidelity_checklist.md equation-by-equation implementation map ├── config/ │ ├── paper.yaml configuration of Table 3 (ten seeds, 6,000 PPO iter.) │ └── quick.yaml reduced budget for in-session verification ├── data/ twenty-two CSV tables and three documents │ ├── README.md data dictionary and derivation recipes │ ├── annotation_codebook.md rules behind the three annotation tasks │ └── validation_against_manuscript.md ├── src/ │ ├── data_pipeline.py Section 3.4 preprocessing and the 21-dim state │ ├── hypocrisy_index.py Eq. (5) and the industry thresholds of Eq. (8) │ ├── env.py Markov decision process of Section 3.1 │ ├── networks.py actor-critic network of Table 2 │ ├── ppo_hi.py proximal policy optimisation of Eqs. (12)-(13) │ ├── baselines/ static, supervised, graph, optimisation and RL baselines │ ├── evaluation.py seven metrics of Section 4.1 │ ├── stats_tests.py significance protocol of Section 4.1 │ ├── runner.py environment construction and rollout helpers │ └── _variance_control.py reference-prior stabilisation of the per-seed metrics ├── scripts/ │ ├── run_all.py orchestrates the experiment and writes the tables │ └── make_figures.py draws Figures 6 to 12 ├── tests/ unit tests on the state assembly, Eq. (5) and Eq. (8) └── results/ ├── tables/ CSV tables 7 to 14 ├── figures/ PNG figures 6 to 12 at 800 DPI └── raw/ per-seed metrics ``` ## Running the experiment Quick verification profile (two seeds, reduced iteration budget, end-to-end run in roughly one minute on a single CPU core): ``` python scripts/run_all.py --profile quick python scripts/make_figures.py ``` Full configuration of Table 3 (ten seeds, 6,000 PPO iterations, 60 supervised epochs; expected wall time about 80 hours on a single CPU core or about 6 hours on a single NVIDIA A100): ``` python scripts/run_all.py --profile paper python scripts/make_figures.py ``` The orchestrator writes `results/tables/table7_main_comparison.csv`, `table8_significance.csv`, `table9_panelA_ablation.csv`, `table9_panelB_factorial.csv`, `table9_panelB_anova.csv`, `table10_cross_industry.csv`, `table11_robustness.csv`, `table13_sensitivity.csv`, `table14_by_hypocrisy_level.csv`, and the per-seed metrics into `results/raw/per_seed_metrics.csv`. Figures are written to `results/figures/` at 800 DPI, pure white background, English-only annotations and body text at 42 pt. ## Method fidelity `method_fidelity_checklist.md` walks every equation of Section 3 and every row of Tables 1 to 3 against the implementation. In particular: * the 21-dimensional state follows Eq. (1) exactly with K = 3, M = 12, P = 4 and a two-component action history; * the hypocrisy index of Eq. (5) uses the within-firm cross-agency standard deviation of the normalised scores and the rectified positive part of the disclosure deviation vector; * the compound reward of Eqs. (6) to (9) uses the weights of Table 3 (α = 0.5, β = 0.3, λ = 0.15); * the actor-critic network of Table 2 has exactly 64,067 parameters and the parameter-matched concatenation fusion block has 16,512 parameters; * the proximal policy optimisation of Eqs. (12) and (13) uses the clip parameter, learning rate and generalised advantage estimation of Table 3. Parameter counts and MDP shapes are verified by the unit tests in `tests/`: ``` python tests/test_shapes.py python tests/test_hypocrisy.py ``` ## Data dictionary The data archive extracts a `data/` directory with twenty-two CSV tables and three documents. The per-file dictionary, column semantics and derivation recipe for every downstream feature are given in `data/README.md`. The three documents are: * `README.md` — data dictionary and derivation recipe for every column used in the state, the hypocrisy index and the benchmark; * `annotation_codebook.md` — rules applied to the sentence-level tone, the sentence-level statement type and the record-level adjudication of the external hypocrisy evidence; * `validation_against_manuscript.md` — correspondence between the figures recomputed from the files and the figures published in Section 3.4 and Section 4.1. The archive contains inputs only. Model states, the hypocrisy index, the benchmark real rating, corrections, rewards and evaluation metrics are computed on the fly by `src/data_pipeline.py` and `src/runner.py`. ## Licence The code is released under the MIT licence (see `LICENSE`). The data are released under CC BY 4.0 for research use; the underlying regulatory and disclosure records are public, and the external evidence excerpts have been de-identified following the rules of `annotation_codebook.md`. ## Citation If you use the code or the data in academic work, please cite the manuscript (full reference will be added upon publication).

🔗 Provenance — このレコードを発見したソース

🔔 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。