ポイント・イン・タイム・プロトコル下でのESG開示とメカニズム型符号付き企業間グラフによる企業グリーンイノベーションの予測

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:ESG開示テキストと企業間グラフを統合した予測モデルとポイント・イン・タイム評価設計は、開示の実質性研究のベンチマークになり得る。

🏢実務担当者:自社のESG開示と取引関係データがグリーンイノベーション評価にどう映るかを理解する手がかりとなる。

🏛政策担当者:開示の実質性やグリーンウォッシングを機械的に検証する枠組み設計の参考になる。

# CMP-DSTCL Forecasting corporate green innovation from ESG disclosures and mechanism-typed signed interfirm graphs under a point-in-time protocol. This repository holds the code and the dataset that accompany the manuscript. Everything runs on a single Python environment and is tested against the ten-seed, two-refit protocol described in the Supplementary Information. ## Repository ``` cmp-dstcl/ Reference implementation (train, baselines, evaluation, tables, figures) cmp-dstcl-data/ Panel, text, graph and audit tables ``` Clone this repository and the two folders will be side by side, which is the layout the configs expect through `paths.dataset: ../cmp-dstcl-data`. ## Quick start ```bash git clone cd cmp-dstcl python -m venv .venv && source .venv/bin/activate pip install -r requirements.txt # Full pipeline under the reported protocol python scripts/run_all.py --config configs/default.yaml --tag full # Short verification on a reduced universe python scripts/run_all.py --config configs/smoke.yaml --tag smoke # Rebuild Tables 3 to 9 and Figures 7 to 12 from the published-protocol run python scripts/make_tables.py --source results/main_run python scripts/make_figures.py --source results/main_run ``` ## What the code implements - Point-in-time protocol with the availability-date rules of Table 1 and the rolling partitions of Equation (2), `src/data/point_in_time.py` - Frozen text encoder (Equation 3) with mean pooling, `src/data/text.py` - Commitment-stripped twin generation (Equation 4) with the acceptance rule of Note S5, `src/data/text.py` - Commitment anchor and projection head (Equations 5 and 6), `src/models/cmp_dstcl.py` - Greenwashing contrast (Equation 7) and greenwashing-gap score (Equation 8), `src/models/losses.py` - Edge-mechanism classifier (Equation 10) with the entry gate of Note S6, `src/models/mechanism.py` - Polarity-aware, mechanism-typed message passing (Equations 11 to 14) with the decay kernel, `src/models/layers.py` - Prior regulariser (Equation 15) and gated recurrent state (Equation 16), `src/models/layers.py`, `src/models/cmp_dstcl.py` - Radial basis outcome encoding (Equation 17) and prognostic contrast (Equations 18 and 19), `src/models/heads.py`, `src/models/losses.py` - Hurdle negative binomial head (Equations 20 and 21), `src/models/heads.py` - Uncertainty weighting (Equation 22), `src/models/losses.py` - Algorithms S1 and S2, `src/models/layers.py`, `src/train.py` - Eighteen baselines of Section 4.1, `src/baselines/models.py` - Metrics, paired bootstrap, DeLong, Holm and FDR control, `src/eval/` - Tables 3 to 9 and Figures 7 to 12 at 800 dpi, `src/eval/tables.py`, `src/figures/` ## What the dataset contains The dataset covers 4,186 A-share listed firms, 58,472 firm-quarters between 2013Q1 and 2023Q4, with annual ESG reports, green patent records, environmental and securities sanctions, five sources of interfirm relations and the three transfer-market windows. Every row carries the date from which it becomes observable. See the dataset README for the file-by-file description. ## Modalities The pipeline reads text, counts and dated records only. There is no audio, video or image input at any stage, and the study involves no survey instrument. The human-coded materials are the edge-mechanism gold standard, its codebook, the green patent taxonomy audit and the blind audit of disclosures. ## Backends `encoder.backend` selects the frozen text encoder: `hf` loads the model named in the configuration, `hashing` is an offline fallback for machines without the weights. `twins.backend` behaves the same way: `vllm` serves the generator named in the configuration, `offline` applies a rule-based rewriter with the same acceptance rule. The protocol, partitions, model and evaluation are identical under either backend. ## Reproducing the published tables and figures `results/main_run/` holds the predictions and summary metric tables of the ten-seed, two-refit, full-universe training run that produced the published tables and figures. `scripts/make_tables.py` and `scripts/make_figures.py` read from that directory by default, so the artifacts match the manuscript. Pass `--source results/run` to build the same artifacts from a local run of the pipeline. ## Citation Please cite the manuscript this repository accompanies. ## License MIT License, see `LICENSE` in each sub-folder.

• openalex https://doi.org/10.5281/zenodo.23513570first seen 2026-10-22 04:58:06 · last seen 2026-11-06 05:01:13