Carbon Transition Risk and Asset Stranding in Agricultural Enterprises

Which agricultural enterprises are most exposed to carbon transition risk — and which would become stranded assets if carbon were priced? This project answers that with a modular Python pipeline that constructs emissions proxies from enterprise-level financial and climate data, prices them under three NGFS-inspired carbon-price scenarios, and flags enterprises whose projected carbon costs exceed projected revenues. Built as a team research project in the CentraleSupélec × ESSEC double-degree curriculum.

Approach

Data. The Agriculture Financial Risk Dataset (Kaggle): 4,981 enterprises × 17 variables — financials (revenue, expenses, loan amount, debt-to-equity), climate exposure (average temperature, rainfall, drought index, flood risk score), and structure (region, enterprise size, quarter).

Pipeline (scripts/run_pipeline.py runs every step end-to-end):

1. Feature engineering — financial ratios (profit margin, cost ratio, debt ratio) and a composite climate stress index: 0.4·drought + 0.4·flood + 0.2·z(temperature).
2. Preprocessing — one-hot encoding of categoricals, standardization of numerics.
3. Emissions proxies — emissions are not observed in the data, so four candidate proxies are built as weighted z-score composites of activity scale (log expenses or log revenue), input-cost intensity, climate stress, and leverage. Their stability is compared via mean absolute deviation and top-decile overlap before fixing a baseline.
4. Models — Ridge regressions (5-fold CV alpha tuning over a log grid, 80/20 holdout) predicting net profit and the baseline emissions proxy.
5. Scenario pricing — carbon cost indices under three carbon prices benchmarked on NGFS Phase 3 (REMIND) scenarios, in USD 2010: Delayed Transition ($10/t), Net Zero 2050 ($110/t), Divergent Net Zero ($300/t). See images/ngfs_bar_plot.png for the benchmark levels.
6. Projection & stranding — 5-year revenue paths (2% growth, NPV at 5% discount), per-scenario future profit and carbon-risk indices, and a stranded flag where projected carbon cost exceeds projected revenue under the severest scenario (Divergent Net Zero).

Scope assumptions (stated up front given data limits): the dataset is enterprise-level, not investor-level; emissions are proxied, not observed; carbon prices are scenario-based, not forecast. The goal is a climate-financial risk pipeline, not precise emissions accounting.

Contents

Path	What it is
scripts/run_pipeline.py	Single entry point — runs the full pipeline
scripts/eda_report.py	EDA entry point — figures, describe table, summary
src/config.py	All paths, scenario prices, and tunable defaults
src/features.py	Financial ratios + climate stress index
src/preprocessing.py	Encoding and scaling
src/proxies.py	Emissions proxy variants + stability metrics
src/models.py	Ridge training with CV alpha tuning
src/carbon.py, src/outputs.py	Scenario carbon costs, revenue projection, stranding analysis
src/eda.py, src/io.py, src/reporting.py	Plots, I/O, markdown/CSV writers
data/AgriRiskFin_Dataset.csv	Raw Kaggle dataset
data/data_cleaned.csv	Preprocessed snapshot
data/final_results_with_stranding_analysis.csv	Full per-enterprise output (66 columns)
outputs/	Figures, tables, and reports from the last run

Results

• Proxy choice matters more than proxy weights. Proxies sharing a scale base agree strongly (expenses-based v1/v3: 70% top-decile overlap; revenue-based v2/v4: 71%), while proxies across bases agree weakly (23–28%). The scale variable, not the weighting scheme, drives who lands in the top risk decile. Full table: outputs/tables/proxy_stability.md.
• Model metrics confirm the dataset's deterministic structure. Both Ridge models reach R² ≈ 1.0 (RMSE ≈ 1e-6): net profit is an accounting identity of revenue and expenses in this synthetic dataset, and the proxy is a linear composite of available features. The models serve as the pipeline's prediction stage, not as evidence of out-of-sample power.
• Stranding under the severest scenario. At $300/t (Divergent Net Zero), every enterprise in the sample flags as stranded. Because risk indices are z-score based (not unit-consistent dollars), the flag is best read as a scenario-severity ordering rather than a calibrated default probability — a limitation the index framing makes explicit.
• Per-enterprise scenario scores (carbon cost, adjusted profit, carbon risk, future indices) are exported to data/final_results_with_stranding_analysis.csv.

Running

Python 3.11. From the repo root:

pip install pandas numpy scikit-learn matplotlib tabulate

python scripts/eda_report.py    # EDA: histograms, correlation heatmap, describe table
python scripts/run_pipeline.py  # full pipeline: proxies, models, scenarios, stranding

Outputs are written to outputs/figures/, outputs/tables/, outputs/reports/, and data/. All defaults (random seed, CV splits, horizon, growth/discount rates, scenario prices) live in src/config.py.

Team

Oscar Caudreliez, with Daniil, Oulaya, Ulysse, and Yuhan — team research project at CentraleSupélec / ESSEC.