words-to-coordinates: simulation data for measuring the gap between disclosed and physical climate risk

Only the asset locations are real. assets/assets.csv comes from the Global Power Plant Database v1.3.0. Hazard fields, disclosure levels, extraction errors and therefore every result in results/ are synthetic. Owner names were used only to group plants into pseudo-firms and have been removed; no row refers to a real company's disclosure or risk.

Data for the preprint Words to Coordinates: An Uncertainty-Aware NLP–Geospatial Framework for Measuring the Gap Between Disclosed and Physical Climate Risk by Thanatchon Chongmankhong (read online · PDF). The paper proposes the Disclosure Gap Index (DGI), which compares a firm's modelled physical exposure with its disclosure intensity and places it in one of four quadrants (aware, silent risk, cheap talk, benign), derives a closed-form relation between geolocation error σ and exposure fidelity for a hazard with correlation length ℓ (Corr = ℓ² / (ℓ² + σ²)), and propagates extraction, geolocation and hazard-model uncertainty by Monte Carlo. Code: https://github.com/thanatchon36/words-to-coordinates; archived on Zenodo: 10.5281/zenodo.24102121.

Independent personal research. Does not represent the views of any employer or institution.

Files

path	content
assets/assets.csv	real locations. The 11,107 power plants of the 1,247 owners with at least three plants in 76 countries: asset, pseudo-firm id firm, ISO3 country, latitude, longitude, capacity_mw, primary_fuel, and weight, the plant's share of its firm's capacity (used to aggregate exposure)
assets/country_points.csv	the point each country is geocoded to in the country-level variant (Natural Earth label point; asset mean where missing)
results/e1_fidelity.csv	E1, synthetic hazards. Spearman correlation between hazard values at true and degraded locations, for assets and for capacity-weighted firm exposure: 7 correlation lengths ℓ (10–1,000 km) × 7 location-error scales σ (1–250 km) plus country-level geocoding × 30 replicates (1,680 rows)
results/e2_quadrant_accuracy.csv	E2. Four-class quadrant accuracy and silent-risk F1 against the true quadrants, for ℓ = 25 and 500 km, five location variants (exact, 5 km, 25 km, 100 km, country) and extraction error rates 0–0.5, 30 replicates (1,800 rows)
results/e2_quadrant_shares.csv	share of firms in each true quadrant per E2 replicate
results/e3_firm_predictions.csv	E3. One row per firm and replicate (20 × 1,247 = 24,940): true exposure rank, true disclosure level (0–4), true quadrant, whether the firm is truly silent risk, the point classification, and P(silent) from Monte Carlo propagation of extraction error only, location + hazard-model error only, and all three (ℓ = 50 km, σ = 25 km, extraction error 0.25, five hazard-model members, 80 draws)
results/e3_posterior_tables.csv	the calibrated P(true level | extracted level) used in each E3 replicate
results/e3_summary.csv	Brier score, expected calibration error, average precision and precision/recall/share flagged at P(silent) ≥ 0.8, mean and s.d. over replicates (Table 3 of the paper)
results/results.json, results/results_extra.json	the summaries the paper's figures, table and text are built from
config.json	experiment parameters and the master seed (20261109)
SHA256SUMS	checksums of every file above

firm in assets/assets.csv and in results/e3_firm_predictions.csv is the same id, so per-firm probabilities can be joined to asset locations and mapped.

How the synthetic parts are generated

Hazard fields are unit-variance stationary Gaussian random fields on the sphere with a squared-exponential kernel in chordal distance, sampled with 1,024 random Fourier features. Firm exposure is the capacity-weighted mean of Φ(z) over the firm's plants, ranked within the sample. Disclosure levels follow a latent-awareness model: a latent score 0.4 · Φ⁻¹(exposure rank) + noise, cut into five levels with shares 15/30/25/20/10%. Extraction error moves a level to an adjacent one with the stated probability. Location error is isotropic Gaussian with per-axis s.d. σ. In E3, the hazard-model ensemble members share the true field with weight √(1 − 0.4²) and add independent fields with weight 0.4. Every experimental unit draws from its own stream derived from the master seed, so make experiments in the code repository reproduces these files exactly.

Licence

CC BY 4.0; see LICENSE. assets/assets.csv is derived from the Global Power Plant Database (World Resources Institute, CC BY 4.0), which must also be credited; Natural Earth is public domain.

Citation

@misc{chongmankhong2026words,
  author = {Chongmankhong, Thanatchon},
  title  = {Words to Coordinates: An Uncertainty-Aware {NLP}--Geospatial Framework for Measuring
            the Gap Between Disclosed and Physical Climate Risk},
  year   = {2026},
  doi    = {10.5281/zenodo.24102121},
  url    = {https://thanatchon36.github.io/words-to-coordinates/paper/},
  note   = {Preprint; simulation results use real asset locations with synthetic hazards and disclosures}
}

Global Power Plant Database: Byers, L., Friedrich, J., Hennig, R., Kressig, A., Li, X., McCormick, C. and Malaguzzi Valeri, L. (2018). A Global Database of Power Plants. World Resources Institute.