GHGbench: A Unified Multi-Entity, Multi-Task Benchmark for Carbon Emission Prediction
arXiv:2605.13743v1 [cs.LG] 13 May 2026
Yifan Duan
Siyuan Zheng Lihuan Li Chao Xue Flora Salim School of Computer Science and Engineering University of New South Wales [email protected] [email protected]
Abstract Open datasets and benchmarks for entity-level carbon-emission prediction remain fragmented across access, scale, granularity, and evaluation. We introduce GHGbench, an open dataset and benchmark for company- and building-level greenhouse-gas prediction. The company track contains 32,000+ company-year records from 12,000+ firms with Scope 1+2 and Scope 3 disclosures and financial/sectoral signals; the building track harmonises 491,591 building-year records from 13 open sources into a single schema across 26 metropolitan areas (10 U.S., 15 Australian, 1 Singaporean), with climate covariates and multimodal remotesensing embeddings. GHGbench defines canonical splits with in-distribution and cross-region/city transfer as primary tasks and temporal hold-out plus short-horizon forecasting as supplementary appendix evidence; headline baselines span gradientboosted trees, a tabular foundation model, MLP, FT-Transformer, and multimodal fusion, with an LLM panel as auxiliary, all evaluated under multi-seed pairedbootstrap tests. Three benchmark-level findings emerge: (i) building emissions are structurally harder than company emissions; (ii) the in-distribution to out-ofdistribution gap dwarfs any within-model gap across both the company track and the building track, and a tabular foundation model is, to our knowledge, the first baseline to open a paired-bootstrap-significant gap over tuned trees on a multi-city building-emissions task; (iii) multimodal remote-sensing embeddings help precisely where tabular generalisation breaks. GHGbench also exposes catastrophic city transfer and the sector-factor lookup ceiling as systematic failure modes. Code and reconstruction recipes are available at GHGbench.
1
Introduction
Global commitments to net-zero greenhouse-gas (GHG) emissions have made quantitative emission prediction a practical requirement for climate policy [23], finance [6], and urban operations [10]. Although emissions accounting has long been a policy concern [41], machine-learning-based entitylevel prediction is still an emerging area [34], and the open benchmarking ecosystem has not caught up [30]. Company-level studies often rely on licensed disclosure or financial feeds [29, 35], while building-level data remain scattered across city portals with incompatible schemas [26, 44]. As a result, it remains difficult to compare models on open, reproducible, entity-level tasks that reflect the geographic, temporal, and schema shifts encountered in real emissions-accounting workflows. Existing approaches cluster into three lines of work, each with structural limits. Company-level studies show that financial, sectoral, and disclosure features can predict Scope 1/2 and Scope 3 emissions, but they often depend on licensed feeds such as CDP, Bloomberg, Refinitiv, or Compustat [21, 29, 30, 35]. Building-energy and operating-carbon benchmarks provide valuable data, but are typically limited by geography, incompatible public schemas, or non-reusable splits [15, 26, 27, 44]. Sector-level emission-factor resources support enrichment, but aggregate above the firms and buildings where Preprint.
Figure 1: GHGbench overview. Left: fragmented emissions resources. Middle: harmonised company + building records with climate and satellite signals. Right: canonical tasks, metrics, and uncertainty protocols. predictions are needed [19, 20, 36]. These limitations motivate a benchmark that evaluates company and building emissions under shared tasks, canonical splits, metrics, and uncertainty reporting. To address this gap, we introduce GHGbench, to our knowledge the first open dataset and benchmark to jointly evaluate company- and building-level carbon-emission prediction under a shared task suite that spans in-distribution regression, cross-region/city transfer, short-horizon forecasting, and multimodal fusion, with paired-bootstrap uncertainty reporting on the headline structured-baseline comparisons. The company track evaluates Scope 1+2 and Scope 3 emissions from corporate disclosures enriched with financial and sectoral signals, released through reconstruction recipes that respect source-data terms. The building track harmonises 491,591 building–year records from 13 open sources across 26 metropolitan areas in the United States, Australia, and Singapore, with unified units, property types, coordinates, and annual GHG targets. Across both tracks, GHGbench provides canonical splits, shared metrics, and uncertainty reporting, while auxiliary weather and satellite-derived signals [14, 28] support multimodal evaluation. Figure 1 summarises this flow. Specifically, the company track starts from public corporate greenhouse-gas disclosures: we normalise identifiers and units, construct Scope 1+2 and Scope 3 panels with deduplication and outlier filtering, and enrich each company-year with financial fields, sectoral emission factors, sub-sector target encoding, and business-summary text for LLM baselines, producing matched wide and strict feature regimes. The building track harmonises 13 public disclosure sources into one schema with a canonical 22-category property taxonomy, per-building-year NASA POWER [28] climate covariates, geocoded coordinates, and Sentinel-2 [14] + Clay multimodal embeddings, yielding 491,591 building-year rows from 100,984 buildings. Both tracks share a multi-task suite spanning in-distribution regression, cross-region/city transfer, temporal hold-out, cross-property-type stress testing, and short-horizon forecasting; we evaluate tuned gradient-boosted trees, TabPFN v2 [22], MLP, FT-Transformer [17], time-series foundation models [1, 11, 40], and an LLM panel under multi-seed runs, bootstrap confidence intervals, and 1000-resample paired bootstrap on shared test rows. Three benchmark-level findings emerge. (i) Building emissions are structurally harder than company emissions: companies are dominated by firm size, while building emissions also depend on occupancy, runtime, and behaviour that public disclosure does not capture. (ii) The in-distribution to out-ofdistribution gap dwarfs any within-model gap on both tracks; on the harder building OOD setting, a pretrained tabular foundation model is, to our knowledge, the first baseline to open a pairedbootstrap-significant gap over tuned trees on a real-world, multi-city, multi-source benchmark. (iii) Multimodal remote-sensing embeddings help precisely where tabular generalisation breaks: a no-op in-distribution, measurable gain on cross-city transfer. The benchmark also exposes systematic failure modes (catastrophic city transfer and the sector-factor lookup ceiling) that mark where deep, multimodal, and foundation models must improve next. Our contributions are: (i) an open dataset and benchmark with 32,000+ company-year records reconstructable from free public APIs and 491,591 harmonised building-year records across 26 metros in three countries; (ii) a multi-task suite with canonical splits for in-distribution, temporal, cross-region/city, cross-property-type, and forecasting evaluation, aligned climate and remote-sensing signals, and a paired-bootstrap protocol; (iii) three benchmark-level findings (cross-entity difficulty 2
asymmetry, ID-to-OOD gap dominating within-model variation, and multimodal embeddings helping where tabular generalisation breaks); and (iv) an open-source release with code, data, and a onecommand quickstart kit.
2
Related Work
Company-level emissions prediction and ESG disclosure. Machine-learning approaches to corporate emissions estimation have largely focused on company-level financial, sectoral, and ESG features. Prior work estimates Scope 1+2 emissions from firm attributes and disclosure-linked financial data [4, 21, 29], extends prediction to Scope 3 categories [35, 42], and documents the large divergence among commercial Scope 3 providers [30]. A related line estimates emissions through sector classification: EXIOBASE input–output factors [36] are packaged as ML-ready benchmarks in ExioML [19], while GREEN/ExioNAICS [20] maps free-text company descriptions to NAICS sectors. Corporate climate text has also become an NLP target, including ClimateBERT [39], GHGextraction benchmarks [5], and open LLM-finetuning corpora [31]. These studies typically isolate one target family, one data source, or one text-extraction task; GHGbench’s company track instead evaluates Scope 1+2, Scope 3, text/LLM baselines, and short-horizon forecasting under one protocol, alongside a building-level benchmark. Building energy and operating-carbon benchmarks. Building-level energy prediction and benchmarking provide the closest analogues for GHGbench’s building track [3, 12]. The Building Data Genome Project 2 [27] releases hourly meter data from 1,636 non-residential buildings, and BuildingsBench [15] adds a large simulated stock with zero-shot/transfer-learning tasks for short-term load forecasting. Closer to annual operating carbon, Li et al. [26] benchmark twelve U.S. cities under a single national schema, while Yap et al. [44] combine disclosures from five cities (New York, Seattle, Washington DC, Melbourne, Singapore) with multimodal geospatial inputs and graph deep learning. Satellite and rooftop imagery have been used as auxiliary signals for building-level energy prediction [13, 37], and geospatial encoders such as SatMAE, Scale-MAE, Prithvi-EO, and Clay [9, 33, 38] make Sentinel-2 [14] embeddings readily usable. Unlike BuildingsBench [15], which targets shortterm load forecasting on simulated meter data, GHGbench targets annual operating-carbon prediction on real disclosures, and differs from prior building-emissions resources along four axes that together define the benchmark gap. (i) Coverage: 26 metros across the U.S., Australia, and Singapore, adding 15 Australian cities under a NABERS/BEEC schema harmonised with U.S. ENERGY STAR Portfolio Manager disclosures. (ii) Evaluation protocol: canonical building-grouped, leave-onecity-out, leave-one-property-type-out, temporal hold-out, and one-step forecasting splits released as the benchmark contract, rather than a single in-distribution regression setting. (iii) Statistical rigour: every headline model-ordering claim is accompanied by a 1000-resample paired bootstrap on shared test rows. (iv) Multimodal scope: aligned NASA POWER climate covariates and Sentinel-2 + Clay embeddings released as a transfer-augmenting signal evaluated separately on in-distribution and cross-city splits. Cross-entity evaluation and modern baselines. Across both entity levels, the central open question is not only which model performs best in distribution, but how model behaviour changes under schema, regional, temporal, property-type, and modality shifts. Tree ensembles remain strong in typical indistribution tabular regimes [18], and TableShift [16] provides a general distribution-shift benchmark for tabular classification. Modern baselines now span TabPFN v2 for tabular prediction [22], Chronos, TimesFM, and Moirai for time-series forecasting [1, 11, 40], and language models for disclosure understanding [20, 31, 39]. GHGbench uses these as evaluation targets, asking how classical, neural, language, time-series, and satellite-augmented models behave when the data are heterogeneous, multimodal, and explicitly split to test generalisation.
3
Dataset
GHGbench is organised around two entity levels that are central to emissions accounting but rarely benchmarked under a shared protocol: companies and buildings. The company track uses annual corporate disclosures to evaluate company-level Scope 1+2 and Scope 3 prediction. The building track harmonises public building-energy and emissions disclosures into an annual building-level operatingcarbon benchmark. Both tracks are built from noisy real-world disclosures and augmented with structured external signals, so models face the data constraints of real emissions-accounting systems. 3
5000 0
5.3k
US
4.3k
EU APAC Other
105
491,591 rows 100,984 buildings
US
AU
SG
104 103 102 NYC L Seatt A le D Chica C go SF Bosto Denv n Po er Phila rtland delph Sy ia Melb dney ou Brisb rne ane Pe Canb rth e Ade rra Gold laide Coas Ho t New bart cast D le Wollo arwin n Town gong sville Cairn s Port Geelong Macq u Singa arie pore
9.8k
10000
Building track building-year rows
company-year rows
Company track 15000 32,830 rows 13.4k 12,087 firms
Figure 2: Dataset coverage. Left: company-year rows by region. Right: building-year rows by metropolitan area, grouped by country. The building track is released as a fully open dataset; the company track ships as a one-command reconstruction script that pulls disclosures via a free public API. 3.1
Company Track
The company track starts from disclosure records collected by the Climate Data Utility (CDU),1 which provides annual company-level greenhouse-gas disclosures from 2018 to 2023. We normalise company identifiers and country codes, convert disclosure fields to numeric tonnes CO2 e, construct Scope 1+2 from reported Scope 1 and location-based Scope 2, deduplicate company–year records, and apply paper-facing outlier filters. The resulting enriched panel contains 32,830 company–year rows from 12,087 companies, with 31,331 usable Scope 1+2 labels. Scope 3 is joined from the raw CDU file at run time, with 18,763 lenient-coverage rows and 9,814 strict-coverage rows after row-level filtering (these are Scope 3 row regimes, distinct from the Scope 1+2 feature regimes defined below). Each company-track record combines country, year, and sector metadata with financial and text signals. Public-company enrichment uses name-to-ticker matching, cached Yahoo-Finance financial statements, company business summaries as text inputs for LLM baselines, and a foreignexchange correction for local-currency financial fields. Sectoral emissions are supplied by an ExioML/EXIOBASE-derived lookup over GICS sector, ExioML region, and reporting year [19, 36], with SICS sub-sector target encoding as an additional sectoral signal; an ExioNAICS/GREEN-style sector-text linkage is provided as a complementary classification resource only [20]. The company track exposes four feature configurations used as the column axis in the headline experiments: Open (country, sector, year, and the ExioML factor on the full panel), Open matched (the same features restricted to the yfinance-matched subset for an apples-to-apples comparison), +firm (Open matched plus revenue, employees, market capitalisation, and EBITDA), and +SICS (+firm with SICS subsector target encoding). Filtering rules are detailed in Appendix C. Because the company track spans 2018–2023, its forecasting task is necessarily short-horizon. 3.2
Building Track
The building track contains 491,591 building–year records from 13 public building-performance sources in the United States, Australia, and Singapore: 100,984 buildings, 26 metro-level cities (10 U.S., 15 Australian, 1 Singaporean), and 471,070 rows with annual operating-GHG targets. The raw year span is 2011–2026, with broad multi-city coverage strongest through 2024. All sources are mapped into one schema for city, building identifier, year, property type, floor area, energy intensity, rating score, GHG emissions, energy use, coordinates, and coordinate provenance. The harmonisation layer standardises units, collapses duplicate building–years, removes or blanks physically invalid values, and maps raw property strings into a 22-category taxonomy. It also resolves source-specific issues that affect model comparability: Australian suburb/state labels are mapped to 13 metro-level cities, NABERS/BEEC annual consumption is converted from MJ to kBtu (1 MJ = 1 https://www.climatedatautility.org/; free registration provides full API access.
4
0.9478 kBtu) before computing EUI, and source-provided coordinates are retained where available. For sources that publish addresses but not coordinates, we run building-level geocoding against Nominatim (OpenStreetMap), rate-limited to 1 request per second with a resumable cache and de-duplicated by query. In the canonical table this supplies 79,155 geocoded building–year rows, corresponding to 20,511 unique buildings across Los Angeles, Boston, Singapore, Denver, and Portland; each match is filtered by a city-radius sanity check before merging with explicit coordinate provenance. Denver, Philadelphia, and Portland have sparse reporting histories and are excluded from temporal and forecasting evaluations while retained in all other experiments. Additional source-specific rules appear in Appendix C. Feature tiers. Because public disclosures expose different fields across countries, building features fall into nine feature tiers organised in three ladders. A 26-city cross-country ladder uses only fields available in every source (size, coordinates, year, degree days). A U.S. ladder progressively adds property type, year built, ENERGY STAR score, source-EUI, and direct energy use as cities permit; an Australian ladder adds site EUI and NABERS-derived rating tiers. Each tier is labelled clean, proxy-rich, or direct-energy-proxy by whether its features include target-correlated proxies, and headline numbers are reported within tier since cross-tier differences reflect richer disclosures rather than model superiority. The full registry is in Appendix H; for cross-type evaluation, the 22-category taxonomy is further collapsed into seven coarse property groups (Appendix C). 3.3
External Alignment
The building track is augmented with building-year climate features from the NASA POWER daily point API [28], aggregated into degree days, annual mean temperature, relative humidity, solar radiation, and wind speed for 491,416 building–year rows. The multimodal extension aligns building-track rows to Sentinel-2 imagery [14] and Clay image embeddings.2 Headline multimodal experiments use 369,698 rows with valid building-footprint patches and target labels; tabular and tabular-plus-Sentinel experiments share this eligible subset. We compare three S2 feature variants: raw 1024-dimensional Clay embeddings, PCA-64, and PCA-128, all fitted on training rows only.
4
Benchmark Design
GHGbench specifies a task suite, evaluation splits, metrics, and baseline families rather than a single leaderboard. This section formalises the evaluation rules used in the experiments; data and feature definitions (including the building feature tiers) are in Section 3. 4.1
Task Suite
GHGbench operationalises four evaluation capabilities rarely tested jointly. These capabilities operationalise the three benchmark-level questions from §1: cross-entity difficulty (in-distribution on both tracks), the ID-to-OOD gap (cross-distribution transfer), and modality lift under shift (multimodal × cross-city). Two primary capabilities (in-distribution prediction, cross-distribution transfer) frame the main results; two supplementary axes (temporal drift, one-step forecasting) appear in the appendix. The full task index is in Appendix Table 4. In-distribution and unseen-entity accuracy. How well do models fit training-distribution data, and how does accuracy hold up on entities unseen during training? Company Scope 1+2 and Scope 3 regression use stratified random splits; building emissions use two split variants—row-random (rows shuffled at row granularity; sensitivity analysis only) and building-grouped (rows partitioned by building_id so every test building is held out from train; the main-paper deployable setting, since real deployments score on entities unseen during training). Cross-distribution transfer. How much do predictions degrade when training and evaluation distributions diverge? At the company level, we leave one of {U.S., EU, APAC} out at a time, training on the other two and evaluating on the third. At the building level, we hold out one of 26 metros at a time; spanning the U.S., Australia, and Singapore, this also exposes country and schema shift. As an appendix-level stress test, we additionally hold out one of seven property categories at a time (the 22-category taxonomy collapsed as described in §3.2). 2 Clay v1.5: https://github.com/Clay-foundation/model.
5
Supplementary capabilities. Temporal drift asks whether models trained on past years generalise forward when reporting cohorts and energy mixes change; the building-level hold-out trains on years ≤ 2019, validates on 2020, and tests on years ≥ 2021 (Appendix Q). Forecasting asks whether lagged context alone suffices for future-period prediction; the company panel forecasts 2022 emissions from 2018–2021 context, while the building panel evaluates each test year y ∈ {2021, 2022, 2023, 2024} separately as horizon h = y−2019 from a fixed train cut at 2019, exposing horizon decay (Appendix Table 16). Both axes use a short five-year panel and overlap with the temporal hold-out, so they are reported as appendix-level evidence (Appendix P). 4.2
Metrics and Reporting Convention
Company-track baselines are evaluated in log10 space, where we report log-space MAE, RMSE, R2 , and Pearson correlation, plus raw-scale percentage errors derived after exponentiating predictions. Building-track baselines are evaluated in raw tonnes CO2 e, where we report MAE, RMSE, MAPE, R2 , NRMSE, and LogMAE. Headline tables and figures use R2 together with MAE or LogMAE, depending on whether the comparison emphasises raw-scale interpretability or heavy-tail robustness; we do not use MAPE as a standalone headline metric because near-zero emissions inflate it disproportionately even with the standard ε-stabilised denominator. 4.3
Statistical Protocol
Every headline number is reported as the mean over three seeds with bootstrap 95% confidence intervals (1000 resamples of the test rows). Pairwise model comparisons on a shared task and feature configuration use a paired bootstrap on the same test rows: for each of 1000 resamples we recompute both models’ losses on the resampled indices and record the per-resample loss difference, yielding a two-sided p-value on ∆ and a 95% CI. We call a gap paired-bootstrap-significant when the resulting CI excludes zero (equivalently p < 0.05). This is the basis for the ID-to-OOD gap and TabPFN-vs-trees claims in §1 and §5. Concrete split parameters (stratification, train/val/test ratios, year cuts, excluded cities) are listed in Appendix E. 4.4
Baselines
Across tasks we evaluate baselines spanning three families: naive references, classical machine learning, and neural networks. The naive references are a global mean, a sector- or city-type group mean, and (company track only) an ExioML/EXIOBASE-derived emission-factor lookup mapping (GICS sector, ExioML region, year) to a tonnes-CO2 e estimate; forecasting tasks additionally use persistence, three-year moving average, linear trend, and a sector-growth extrapolation. Classical machine-learning baselines are Ridge regression, RandomForest [7], XGBoost [8], and LightGBM [25], with a lag-feature gradient-boosted regressor reused for forecasting. Neural baselines comprise a multi-layer perceptron and a panel of pretrained foundation models: TabPFN v2 [22] for tabular regression; Chronos [1], TimesFM [11], and Moirai [40] for time-series forecasting; and an LLM panel of Claude [2], GPT [32], Qwen [43], and Mistral [24] on the company track.
5
Experiments and Analysis
We report headline results on both tracks under the protocol in §4.3, organised around three benchmark-level findings: a cross-entity difficulty asymmetry (§5.1), an ID-to-OOD gap that dominates within-model variation (§5.2), and a multimodal lift concentrated on cross-city transfer (§5.2). Single-seed MLP numbers are flagged where they appear and are not used for ordering claims. 5.1
Main Results
Company track. Company-track performance is governed primarily by information regime, not by small differences between strong models. (i) With Open features alone the eight learned baselines cluster at R2 ≈ 0.28–0.31 (Table 1), barely above the sector-mean floor (0.232 on Open, 0.147 on the matched panel)—country, year, and the ExioML factor add little beyond sector. (ii) The yfinancematched subset drops R2 by 0.03 (TabPFN v2) to 0.11 (XGBoost), confirming the matched panel is genuinely harder; TabPFN v2 leads on Open matched (0.268), consistent with its low-data prior 6
Table 1: Company-track regression, R2 (computed in log10 space; feature regimes defined in §3.1). Rows are grouped by method family: naive (Sector mean), classical ML (Ridge, tree ensembles), and deep / foundation (MLP, FT-Transformer, TabPFN v2). Cross-region: leave-one-region-out over US/EU/APAC, mean on +firm. Bold: column max. Scope 1+2
Model
Scope 3
Cross-region
Open Open matched +firm +SICS Open Open matched +firm +SICS
+firm
Sector mean
0.232
0.147
0.147
0.147
0.162
0.064
0.064
0.064
0.055
Ridge RandomForest LightGBM XGBoost HistGradientBoosting
0.284 0.288 0.278 0.282 0.300
0.229 0.201 0.180 0.177 0.235
0.549 0.875 0.873 0.879 0.871
0.613 0.885 0.880 0.885 0.869
0.211 0.181 0.190 0.213 0.215
0.128 0.059 0.048 0.072 0.059
0.370 0.639 0.637 0.645 0.653
0.411 0.687 0.688 0.681 0.677
0.506 0.485 0.478 0.468 0.452
MLP FT-Transformer TabPFN v2
0.312 0.313 0.297
0.265 0.246 0.268
0.693 0.622 0.857
0.747 0.690 0.873
0.231 0.212 0.235
0.127 0.129 0.139
0.460 0.418 0.633
0.497 0.461 0.631
0.414 0.494 0.467
advantage. (iii) Adding firm-level features lifts the strongest tree and TabPFN models to R2 ≈ 0.86– 0.88—a +0.59 to +0.70 jump, an order of magnitude larger than the within-tree gap (≤ 0.008); SICS sub-sector encoding adds a further ∼ +0.01 on +firm, a smaller but consistent second-tier lever. On +firm TabPFN v2 is paired-bootstrap-tied with every tuned tree (|∆R2 | ≤ 0.017, p ≥ 0.14), with trees partly distinguishable on shared rows (Appendix S). (iv) Scope 3 follows the same regime pattern at a lower absolute level (best R2 = 0.688, LightGBM+SICS); cross-region transfer is the only company-track setting where Ridge (0.506±0.053) beats the tuned trees (cross-region means 0.452– 0.485), consistent with linear models being less prone to overfit region-specific interactions. Auxiliary forecasting and LLM panels (Appendix Tables 14, 13) show simple temporal priors dominate annual forecasting and LLMs trail strong models by 0.32–1.12 R2 on the n = 200 overlap. Building track. Table 2 reports the core building regression setting. (i) Building prediction is intrinsically harder than company prediction: the strongest building model (TabPFN v2 at R2 = 0.479 on building-grouped) sits well below the strongest company model (R2 = 0.879, Table 1). The two tracks use different target scales, but the gap reflects a genuine task-level difference—company emissions are dominated by firm size (revenue alone is the top single predictor, Appendix L), while annual building emissions also depend on occupancy schedules, runtime, equipment efficiency, and occupant behaviour, factors that no public disclosure captures and vary widely between similar buildings; the benchmark deliberately spans this heterogeneity (26 metros, three countries, 13 portals), and the lower attainable R2 is a property of the task. (ii) On unseen buildings, TabPFN v2 is, to our knowledge, the first baseline to open a paired-bootstrap-significant gap over tuned trees on a multi-city, multi-source building-emissions panel. Under the building-grouped split (every test building held out from train—the deployable setting), TabPFN v2 reaches R2 =0.482±0.029, beating LightGBM by +0.025 (p = .012) and XGBoost by +0.036 (p = .002) on shared test rows, and tying with RandomForest (+0.010, p = .36); adding the NASA POWER climate channels makes the gap significant against all three trees (p ≤ .001, Appendix S). The ranking inverts under the row-random split, where RF leads (0.477 vs. TabPFN 0.405)—confirming that the gap is specific to the unseenentity regime and not an artefact of any tabular property of the data. (iii) Switching to building-grouped drops RF by ∼ 6 pp (0.477 → 0.415) and trims tabular MLP modestly (0.353±0.08 → 0.341±0.08); adding the four NASA POWER climate channels yields a similar grouped MLP at 0.346±0.06. Tree splits on raw lat/lon already partition the space along climate-correlated boundaries, whereas MLP’s dense first layer must learn this geographic prior end-to-end and trails the tuned trees by ∼ 0.07 R2 on the grouped split. FT-Transformer, despite its richer transformer architecture, also reaches only R2 = 0.403±0.023 on Core grouped, sitting between MLP and the tuned trees and never matching TabPFN. This confirms that the foundation-model edge comes from in-context pretraining rather than from transformer architecture alone. Multimodal S2 results, cross-tier and foundation-model rankings, and analyses of schema and transfer effects are discussed in Section 5.2 onwards; temporal hold-out (Appendix Q), cross-property-type stress (Appendix R), and forecasting (Appendix P) are reported as appendix-level evidence. 7
Table 2: Building-track regression on the 26-city cross-country core panel (§3.2). +S2: trees/TabPFN use concat fusion; MLP uses residual fusion (§5.2). Rows grouped by method family: naive, classical ML, deep / foundation. Bold: column max. Row-random
Model Core
Building-grouped
+ Climate
+ S2
Core
+ Climate
−0.045 ± 0.002 −0.045 ± 0.002 0.031 ± 0.002 0.031 ± 0.002
Ridge RandomForest LightGBM XGBoost
−3.47 ± 1.81 0.477 ± 0.030 0.413 ± 0.020 0.410 ± 0.020
−3.36 ± 1.83 0.476 ± 0.028 0.414 ± 0.020 0.415 ± 0.023
−3.98 ± 2.11 0.444 ± 0.012 0.401 ± 0.017 0.401 ± 0.017
−3.86 ± 1.47 0.415 ± 0.030 0.401 ± 0.023 0.399 ± 0.025
−3.75 ± 1.41 0.413 ± 0.028 0.401 ± 0.026 0.399 ± 0.026
−3.08 ± 2.01 0.427 ± 0.021 0.399 ± 0.012 0.397 ± 0.013
MLP FT-Transformer TabPFN v2
0.353 ± 0.076 0.417 ± 0.039 0.405 ± 0.071
0.343 ± 0.080 0.394 ± 0.018 0.409 ± 0.067
0.334 ± 0.150 0.408 ± 0.017 0.377 ± 0.065
0.341 ± 0.078 0.403 ± 0.023 0.482 ± 0.029
0.346 ± 0.059 0.398 ± 0.023 0.482 ± 0.027
0.385 ± 0.078 0.414 ± 0.013 0.380 ± 0.060
— —
Cross-country
US-only
AU-only
Core +climate
Core +metadata +EUI +energy
Core +NABERS +energy
0.8 0.6 0.4
proxy-rich (EUI/NABERS) +direct energy
Grouped-building R 2 (5-seed mean)
GlobalMean CityTypeMean
— —
−0.045 ± 0.004 −0.045 ± 0.004 0.030 ± 0.005 0.030 ± 0.005
+ S2
0.2 0.0
CityTypeMean (floor)
TabPFN v2
RF
LightGBM
XGBoost
MLP
FT-T
Figure 3: Building-track R2 on the building-grouped split across the nine feature tiers and three ladders defined in §3.2 (full registry in Appendix H). Shaded bands mark proxy-rich and directenergy-proxy tiers.
5.2
Analysis and Findings
Sector-factor estimation trails learned models. Predicting emissions by multiplying revenue with the ExioML/EXIOBASE sectoral factor reaches R2 = 0.222 on the firm-matched company panel, well below the tuned tree family (≈ 0.87); a sector-mean predictor sits even lower at 0.147. Sector identity is still informative when used as one feature among many—adding SICS sub-sector target encoding lifts most baselines (Table 1), and tuned-LightGBM attribution ranks the ExioML factor second after revenue—but it is too coarse to drive prediction on its own. Per-feature attribution and a Financial Services case study with a SICS-granularity check are in Appendix L and Appendix O.1. A pretrained tabular foundation model is consistently competitive and often best. On the building track, a 1,000-resample paired bootstrap on shared test rows (Appendix S, Table 20) shows TabPFN v2 paired-bootstrap-significantly outperforms every tuned tree (p < .05) on three of six building-grouped tiers—26-city core+climate, U.S. core, and U.S. metadata. On the 26-city core tier, TabPFN beats LightGBM (p = .012) and XGBoost (p = .002) but ties RF (p = .36); on the small Australia-only panel all four models are tied. On the company track TabPFN is also tied with the tuned trees (R2 = 0.857, p ≥ 0.14). MLP, FT-Transformer, and LLM baselines remain weaker, so the advantage is specific to TabPFN’s in-context training rather than a blanket win for deep tabular models [18]. 8
US
AU
SG
pooled grouped RF ≈ 0.41
0.4 0.2 0.0 −0.2 Model RF
n
ph
go n
g
ia
XGB
lo
el
P
h
il
ad
ga in S
LGBM
W ol
e
re po
rn
de ai
bo u
de l
el
A
n to os B
LA
a
ey yd n
be an
S
rr
o
C
ag C
M
G
ol
C
d
h
ic
st C
D
in
oa
w
ve r
ar D
en D
le
YC N
d
il n sv To w
n e
an tl P or
th
sb a ri B
S
P er
rn
st l
ai C
N
ew ca
F
Omitted from this figure (RF R 2 < −0.4): Seattle, Hobart. Full per-city table in Appendix.
e
−0.4
s
City-LOCO R 2 (3-seed mean)
Country
0.6
Figure 4: Building-track leave-one-city-out on the 26-city cross-country core tier. Cities sorted by RF R2 descending; dashed line is the pooled-panel Building-grouped RF baseline. Off-scale tree points (Seattle, Hobart) are flagged at the left edge. Per-city table for all models in Appendix 19. Feature quality drives rankings more than model choice. Tuned-LightGBM permutation attribution on the 26-city core tier (Appendix 9) ranks gross floor area first, location second, and climate/time fields third; in Figure 3, moving from clean physical attributes to EUI/rating/direct-energy proxies changes the attainable R2 far more than swapping among strong tree variants. Cross-city transfer is the hardest core building setting. Figure 4 reports three-seed leave-onecity-out results on the 26-city cross-country core tier. Mean-city tree-family R2 collapses from the pooled-panel Building-grouped value of ∼ 0.41 (Table 2) to R2 = 0.03 (LightGBM), 0.08 (XGBoost), and 0.13 (RandomForest). The pattern is heterogeneous rather than a uniform failure: Newcastle, Cairns, and Perth remain at R2 ≈ 0.40–0.69 across all three tuned trees, while Seattle and Hobart turn strongly negative (RF −1.23 on Seattle; LGBM/XGB collapse to below −2 on Hobart). The three tree families largely agree on which cities are easy versus hard, indicating the failures are driven by city-level distribution shift rather than a single model’s brittleness. The full per-city table for all models, including the off-scale tabular MLP failures, is in Appendix Table 19. Sentinel-2 helps cross-city transfer, not in-distribution prediction. On the S2-eligible subset, PCA64 Sentinel-2 + Clay embeddings are nearly a no-op in-distribution (∆R2 ∈[−0.012, −0.007] for trees on the grouped split) because trees already capture geography through lat/lon. On cross-city leaveone-out, mean-city R2 rises by +0.071 (LightGBM, +0.054 → +0.125) and +0.038 (XGBoost), with 11/23 and 14/23 city win-rates respectively; raw lat/lon are unique tokens that fail to transfer, while a Sentinel-2 patch’s visual content (rooftop type, density, vegetation) supplies a generalisable geographic prior. We position Sentinel-2 + Clay embeddings as a transfer-augmenting signal, not a uniform improver—useful precisely where tabular generalisation breaks. The cross-city S2 effect is heterogeneous. San Francisco, Washington DC, and Melbourne gain on all three models, with Boston and Perth additionally gaining on both tree models, while Chicago loses on both trees. S2 is strongest where dense urban form yields a transferable visual prior, weakest when the training panel is dominated by similar-schema cities. A fusion-design ablation and per-city deltas are in Appendix T. Forecasting reveals a panel-length asymmetry across tracks. On the company track’s six-year panel, naive temporal references (sector-growth and linear trend) reach R2 ≈ 0.94–0.96, while TimesFM and Chronos trail at 0.81/0.78 (Appendix Table 14). On the building track’s 5-year panel, time-series foundation models lead at every horizon h ∈ {2, . . . , 5} (TimesFM peaks at R2 = 0.90 on h = 4), with tuned trees within ∼ 0.1 R2 and naive linear trend lagging by ∼ 0.2 (Appendix Table 16). Building forecasting therefore rewards larger temporal models, while the shorter company panel leaves simple temporal priors hard to beat.
6
Conclusion
GHGbench’s primary contribution is a harmonised cross-country GHG dataset with canonical indistribution and cross-region/city splits and a paired-bootstrap protocol. Three findings emerge: a cross-entity difficulty asymmetry, an ID-to-OOD gap that dominates within-model variation, and a multimodal lift concentrated where tabular generalisation breaks. Longer-horizon forecasting, frontier LLMs, and richer multimodal fusion are open follow-ups the benchmark is designed to support. 9
References [1] Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of time series. Transactions on Machine Learning Research, 2024. [2] Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku. https://www.anthropic. com/news/claude-3-family, 2024. [3] Pandarasamy Arjunan, Kameshwar Poolla, and Clayton Miller. EnergyStar++: Towards more accurate and explanatory building energy benchmarking. Applied Energy, 276:115413, 2020. doi: 10.1016/j.apenergy.2020.115413. [4] Jérémi Assael, Thibaut Heurtebize, Laurent Carlier, and François Soupé. Greenhouse gases emissions: Estimating corporate non-reported emissions using interpretable machine learning. Sustainability, 15(4):3391, 2023. doi: 10.3390/su15043391. [5] Jacob Beck, Anna Steinberg, Andreas Dimmelmeier, et al. Addressing data gaps in sustainability reporting: A benchmark dataset for greenhouse gas emission extraction. Scientific Data, 12: 1497, 2025. doi: 10.1038/s41597-025-05664-8. [6] Patrick Bolton and Marcin Kacperczyk. Do investors care about carbon risk? Journal of Financial Economics, 142(2):517–549, 2021. doi: 10.1016/j.jfineco.2021.05.008. [7] Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001. doi: 10.1023/A: 1010933404324. [8] Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016. doi: 10.1145/2939672.2939785. [9] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In Advances in Neural Information Processing Systems, volume 35, 2022. [10] Felix Creutzig, Giovanni Baiocchi, Robert Bierkandt, Peter-Paul Pichler, and Karen C. Seto. Global typology of urban energy use and potentials for an urbanization mitigation wedge. Proceedings of the National Academy of Sciences, 112(20):6283–6288, 2015. doi: 10.1073/ pnas.1315545112. [11] Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. In International Conference on Machine Learning (ICML), 2024. [12] Hengfang Deng, David Fannon, and Matthew J. Eckelman. Predictive modeling for US commercial building energy use: A comparison of existing statistical and machine learning algorithms using CBECS microdata. Energy and Buildings, 163:34–43, 2018. doi: 10.1016/j. enbuild.2017.12.031. [13] Thomas R. Dougherty, Tianyuan Huang, Yirong Chen, Rishee K. Jain, and Ram Rajagopal. SCHMEAR: Scalable construction of holistic models for energy analysis from rooftops. In Proceedings of the 8th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys), 2021. [14] M. Drusch, U. Del Bello, S. Carlier, O. Colin, V. Fernandez, F. Gascon, B. Hoersch, C. Isola, P. Laberinti, P. Martimort, A. Meygret, F. Spoto, O. Sy, F. Marchese, and P. Bargellini. Sentinel2: ESA’s optical high-resolution mission for GMES operational services. Remote Sensing of Environment, 120:25–36, 2012. doi: 10.1016/j.rse.2011.11.026. [15] Patrick Emami, Abhijeet Sahu, and Peter Graf. BuildingsBench: A large-scale dataset of 900K buildings and benchmark for short-term load forecasting. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2023. 10
[16] Josh Gardner, Zoran Popovic, and Ludwig Schmidt. Benchmarking distribution shift in tabular data with TableShift. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2023. [17] Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data. In Advances in Neural Information Processing Systems (NeurIPS), 2021. [18] Léo Grinsztajn, Edouard Oyallon, and Gaël Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. [19] Yanming Guo, Charles Guan, and Jin Ma. ExioML: Eco-economic dataset for machine learning in global sectoral sustainability. In Tackling Climate Change with Machine Learning Workshop at ICLR, 2024. arXiv:2406.09046. [20] Yanming Guo, Xiao Qian, Kevin Credit, and Jin Ma. Group reasoning emission estimation networks. arXiv preprint arXiv:2502.06874, 2025. Tackling Climate Change with Machine Learning Workshop at ICLR 2025; introduces the ExioNAICS dataset. [21] You Han, Achintya Gopal, Liwen Ouyang, and Aaron Key. Estimation of corporate greenhouse gas emissions via machine learning. In Tackling Climate Change with Machine Learning Workshop at ICML, 2021. arXiv:2109.04318. [22] Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model. Nature, 637:319–326, 2025. doi: 10.1038/s41586-024-08328-6. [23] IPCC. Climate change 2023: Synthesis report. contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Technical report, Intergovernmental Panel on Climate Change, Geneva, Switzerland, 2023. [24] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023. [25] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems (NeurIPS), 2017. [26] Tian Li, Tianqi Liu, Azadeh Omidfar Sawyer, Pingbo Tang, Vivian Loftness, Yi Lu, and Jiarong Xie. Generalized building energy and carbon emissions benchmarking with post-prediction analysis. Developments in the Built Environment, 17:100320, 2024. doi: 10.1016/j.dibe.2024. 100320. [27] Clayton Miller, Anjukan Kathirgamanathan, Bianca Picchetti, Pandarasamy Arjunan, June Young Park, Zoltan Nagy, Paul Raftery, Brodie W. Hobson, Zixiao Shi, and Forrest Meggers. The building data genome project 2, energy meter data from the ASHRAE great energy predictor III competition. Scientific Data, 7(1):368, 2020. doi: 10.1038/s41597-020-00712-x. [28] NASA Langley Research Center POWER Project. NASA POWER daily api. https://power. larc.nasa.gov/docs/services/api/temporal/daily/, 2026. Accessed 2026-04-25. [29] Quyen Nguyen, Ivan Diaz-Rainey, and Duminda Kuruppuarachchi. Predicting corporate carbon footprints for climate finance risk analyses: A machine learning approach. Energy Economics, 95:105129, 2021. doi: 10.1016/j.eneco.2021.105129. [30] Quyen Nguyen, Ivan Diaz-Rainey, Adam Kitto, Ben I. McNeil, Nicholas A. Pittman, and Renzhu Zhang. Scope 3 emissions: Data quality and machine learning prediction accuracy. PLOS Climate, 2(11):e0000208, 2023. doi: 10.1371/journal.pclm.0000208. [31] nopperl. Corporate emission reports: LLM finetuning dataset for extracting Scope 1/2/3 from sustainability reports. https://github.com/nopperl/corporate_emission_reports, 2024. Accessed 2026-04-26. 11
[32] OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. [33] Colorado J. Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uyttendaele, and Trevor Darrell. Scale-MAE: A scaleaware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. [34] David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, Andrew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna WaldmanBrown, Alexandra Sasha Luccioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio. Tackling climate change with machine learning. ACM Computing Surveys, 55(2):1–96, 2022. doi: 10.1145/3485128. [35] George Serafeim and Gladys Vélez Caicedo. Machine learning models for prediction of Scope 3 carbon emissions. Technical Report 22-080, Harvard Business School, 2022. URL https://www.hbs.edu/faculty/Pages/item.aspx?num=62566. [36] Konstantin Stadler, Richard Wood, Tatyana Bulavskaya, Carl-Johan Södersten, Moana Simas, Sarah Schmidt, Arkaitz Usubiaga, José Acosta-Fernández, Jeroen Kuenen, Martin Bruckner, Stefan Giljum, Stephan Lutter, Stefano Merciai, Jannick H. Schmidt, Michaela C. Theurl, Christoph Plutzar, Thomas Kastner, Nina Eisenmenger, Karl-Heinz Erb, Arjan de Koning, and Arnold Tukker. EXIOBASE 3: Developing a time series of detailed environmentally extended multi-regional input-output tables. Journal of Industrial Ecology, 22(3):502–515, 2018. doi: 10.1111/jiec.12715. [37] Artem Streltsov, Jordan M. Malof, Bohao Huang, and Kyle Bradbury. Estimating residential building energy consumption using overhead imagery. Applied Energy, 280:116018, 2020. doi: 10.1016/j.apenergy.2020.116018. [38] Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, et al. Prithvi-EO-2.0: A versatile multi-temporal foundation model for Earth observation applications. arXiv preprint arXiv:2412.02732, 2024. [39] Nicolas Webersinke, Mathias Kraus, Julia Anna Bingler, and Markus Leippold. ClimateBert: A pretrained language model for climate-related text. arXiv preprint arXiv:2110.12010, 2021. [40] Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. In International Conference on Machine Learning (ICML), 2024. [41] World Resources Institute and World Business Council for Sustainable Development. The GHG protocol: A corporate accounting and reporting standard (revised edition). https: //ghgprotocol.org/corporate-standard, 2004. [42] World Resources Institute and World Business Council for Sustainable Development. Corporate value chain (Scope 3) accounting and reporting standard. Technical report, Greenhouse Gas Protocol, 2011. URL https://ghgprotocol.org/ corporate-value-chain-scope-3-standard. [43] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. [44] Winston Yap, Angela N. Wu, Clayton Miller, et al. Revealing building operating carbon dynamics for multiple cities. Nature Sustainability, 8:1199–1210, 2025. doi: 10.1038/ s41893-025-01615-8.
12
Table 3: GHGbench dataset overview. Track
Scale
Targets
Key signals
Company
12,087 companies; 32,830 enriched company–year rows from 2018–2023; 31,331 usable Scope 1+2 rows; 18,763 filtered Scope 3 rows.
Annual Scope 1+2 and Scope 3 emissions.
Country, sector, financial features, ExioML/EXIOBASE-derived sectoral factors, and disclosure text/business summaries.
Building
100,984 buildings; 491,591 building– year rows from 13 open sources across 26 metros in the United States, Australia, and Singapore; reporting years 2011– 2026, dense through 2024.
Annual building operating GHG emissions in metric tonnes CO2 e.
Harmonised building attributes, property taxonomy, coordinates, NASA POWER building-year climate, and Sentinel-2/Clay image embeddings.
A
Datasheet for GHGbench
Motivation. GHGbench supports machine-learning research on greenhouse-gas (GHG) emission prediction for two complementary entity types: companies and buildings. The two tracks were assembled because no public benchmark previously evaluated deep, foundation, and multimodal models on heterogeneous, multi-source emissions data under realistic generalisation regimes. Composition. The company track contains 32,830 enriched company–year rows from 12,087 companies (2018–2023), with 31,331 usable Scope 1+2 labels and 18,763 lenient-coverage Scope 3 labels after the filters described in Section 3.1. The building track contains 491,591 building–year rows across 100,984 buildings and 26 metropolitan areas in the United States, Australia, and Singapore, of which 471,070 carry non-null annual GHG labels. Each row is augmented with NASA POWER climate features and optionally with Sentinel-2 + Clay image embeddings (369,698 rows have a valid building-footprint patch in the headline multimodal subset). Collection process. Company labels and disclosures come from CDU exports; financial enrichment is built from yfinance and matched to a ticker. Building disclosures come from 13 city- or countrylevel open-data portals; raw fields are normalised through the harmonisation scripts described in Section 3.2, including a NABERS→kBTU unit fix and a property-type taxonomy of 22 categories. Recommended uses. The benchmark is designed for in-distribution regression, temporal hold-out, cross-region and cross-city transfer, cross-property-type stress testing, short-horizon forecasting, and multimodal alignment. We discourage using the proxy-rich and direct-energy-proxy tiers as deployable headline results; Section 3.2 formalises the reporting rule. Maintenance. Source data are not redistributed in raw form. The release is a reconstruction recipe: code, canonical splits, harmonisation metadata, and download instructions for each portal. Updates will follow source-data releases (annual disclosure cycles) and will be versioned. Table 3 summarises the two tracks in compact form.
B
Limitations and Broader Impact
Limitations. GHGbench’s headline contribution is the harmonised dataset, canonical splits, and paired-bootstrap protocol on in-distribution and cross-region/city transfer; forecasting, LLM-only emission prediction, and advanced multimodal fusion are reported as supplementary axes rather than headline claims, and the limitations below are scoped accordingly. The public-data design also sets limits. The company track depends on CDU self-reported disclosures and is released as a reconstruction recipe rather than a raw-data mirror. The building track covers 26 metropolitan areas and 13 sources, but coverage is skewed toward cities that publish benchmarking data; some cities are temporally sparse, and Los Angeles and Denver rely heavily on address-level geocoding. The Sentinel-2 extension uses only valid building-footprint patches in headline multimodal experiments, and the model panel is baseline-oriented rather than an exhaustive architecture search. Annual forecasting is evaluated only one step ahead on a short five-year panel, so longer-horizon time-series foundation-model behaviour is undersampled. Multimodal fusion uses only late concatenation of PCA-compressed Sentinel-2 embeddings, and LLM-only emission prediction is evaluated as zero/fewshot point regression; we leave more advanced multimodal fusion architectures and reasoning-style LLM pipelines as open follow-up work. 13
Broader impact. GHGbench is intended to support open, comparable evaluation of carbon-emission prediction for climate-policy and urban-operations research that previously depended on paywalled disclosure feeds. As with any predictive benchmark, point-estimate predictions can convey false confidence on noisy panels; we encourage users to consult the per-cell standard deviations and bootstrap confidence intervals we report rather than headline means, especially on small or out-ofdistribution panels (e.g., the AU NABERS panel) where seed variance is high.
C
Data Harmonisation Details
Company preprocessing. The company track converts dash placeholders and disclosure fields to numeric tonnes CO2 e, constructs Scope 1+2 from reported Scope 1 and location-based Scope 2, normalises company identifiers and country codes, and deduplicates company–year records before splitting. For Scope 1+2, labels in [0, 10) tonnes CO2 e or above 5 × 108 tonnes CO2 e are removed as unit or reporting errors. Scope 3 is joined from the raw CDU disclosure file at run time and filtered to records with non-null country and Scope 3 between 100 and 2 × 109 tonnes CO2 e. The lenient-coverage panel uses country, reporting year, and sector metadata; the strict-coverage panel additionally requires a ticker match and positive revenue, and adds time-aligned or fallback financial variables. Building preprocessing. The building track uses source-specific parsers for CSV, XLSX, and XLSB portals; converts floor area to square feet, EUI to kBtu/ft2 , gas use to kBtu, electricity to kWh, and GHG labels to metric tonnes CO2 e; and collapses duplicate building–years using numeric medians and categorical modes. Physically invalid values, including negative EUI/GHG values and extreme EUI outliers, are removed or blanked while preserving rows with other usable fields. The propertytype mapping keeps coarse source labels explicit: for example, San Francisco’s raw “Commercial” label maps to a dedicated mixed-commercial bucket rather than retail. The coord_source column separates source coordinates from address-level geocodes.
D
Task-suite Details
Table 4 gives the full task definition table referenced in Section 4.1.
E
Splits and Generalization Axes
This section gives the concrete split parameters for the four evaluation capabilities defined in Section 4.1. In-distribution splits. Company regression uses a deterministic 80/10/10 train/validation/test split stratified over (GICS sector, country) with rare-stratum fallback, reported on both the lenient-coverage and strict-coverage panels defined in Section 3.1. The lenient-coverage panel illustrates how far broad-coverage metadata alone can go, while the strict-coverage panel carries the headline structured, LLM, and paired-bootstrap comparisons because those require matched financial and text features. The Scope 3 task applies the row filter from Section 3.1 and rebuilds an 80/10/10 stratified split on the filtered panel rather than reusing the Scope 1+2 split. Building regression uses a 70/10/20 split in two variants: a row-random split for sensitivity analyses, and a grouped-building split that keeps every building in a single fold. The grouped variant is the main-paper deployable setting because evaluation buildings are unseen during training. Cross-distribution splits. Company cross-region transfer leaves one of {U.S., EU, APAC} out at a time and is evaluated on the strict-coverage panel. Building cross-city transfer holds out one entire metro at a time, splitting source-city buildings between training and validation while the held-out city contributes only test rows. Building cross-property-type transfer holds out one of seven property categories at a time, reported as appendix-level evidence. Temporal split. The building-level temporal hold-out trains on years ≤ 2019, validates on 2020, and tests on years ≥ 2021; Denver, Philadelphia, and Portland have only one or three reporting years (Section 3.2) and are excluded. The company panel covers only 2018–2023, so company-level temporal generalization is evaluated through forecasting rather than a separate hold-out. 14
Table 4: GHGbench task suite. Track
Task
Target
Train/test rule
Evaluation focus
T1
A
Scope 1+2 emissions
T1
A′
Scope 3 emissions
In-distribution company regression and value of matched financial/text features. Harder indirect-emissions target under the same company protocol.
T1
C
Scope 1+2 emissions
T1
E
Scope 1+2 forecasts
Stratified company–year split, 80/10/10 train/validation/test, stratified by GICS sector and country with rare-stratum fallback; reported on T1-Wide and T1-Strict. Scope 3 is joined from raw CDU disclosures, filtered as in Section 3.1, then rebuilt as an 80/10/10 stratified split with the same wide/strict definitions as T1-A. Leave-one-region-out on T1-Strict: train on two of US, EU, APAC and evaluate on the held-out region. Balanced 2018–2022 company panel; use 2018–2021 context to forecast 2022, with classical, ML, and time-series foundationmodel baselines.
T2
A
Building-year GHG emissions
Feature-tier ladder, targetproxy interpretation, and baseline model ranking.
T2
B
Building-year GHG emissions
T2
C1
Building-year GHG emissions
T2
D
Building-year GHG emissions
T2
E
One-step buildingyear GHG forecasts
T2
S2
Building-year GHG emissions
Pooled and per-city prediction with either row-random or grouped-building 70/10/20 train/validation/test splits; the grouped split is the main-paper deployable setting because each building appears in only one fold. Temporal hold-out: train on years ≤2019, validate on 2020, test on years ≥2021; COVID variant trains on years ≤2018 and evaluates 2019–2021 separately. Leave-one-city-out: train on all source cities, carve validation from source buildings, and test on the held-out city. Leave-one-property-type-out on U.S. tiers with property-type metadata; main stress-test types are Office, Multifamily, Retail, Hotel, K-12 School, Hospital/Medical, and Warehouse/Distribution. Eligible buildings must have enough history and data on both sides of the time cut; train on years ≤2019, validate on 2020, and test on years ≥2021 using lagged emissions and lagged climate only. Re-run T2-A and T2-C1 on the identical S2eligible subset for tabular-only and tabularplus-Sentinel variants; S2 uses raw 1024dimensional embeddings or PCA-64/PCA128 fitted on train embeddings only.
Cross-region corporate generalization. Short annual-panel forecasting and persistence-style baselines.
Temporal robustness under public-disclosure drift.
Cross-city transfer and country/schema heterogeneity. Cross-type stress test; appendix-level supporting evidence.
Short-horizon forecasting without current-year climate information.
Multimodal extension under matched row availability.
Forecasting splits. Company forecasting uses a balanced 2018–2022 panel and predicts 2022 emissions from 2018–2021 context. Building forecasting uses the same year cuts as the temporal hold-out but restricts inputs to lagged emissions and lagged climate, preventing current-year climate leakage.
F
Property-type Taxonomy
Table 5 lists the 22 canonical building property-type categories used in the building track, the typical raw forms each category absorbs, and the cities for which the category is populated. The full mapping JSON is available at scripts/property_type_mapping.json, and the heuristic fallback used when raw values do not match the mapping is in scripts/standardize_buildings.py.
G
Feature Availability across Cities and Tiers
Figure 5 reproduces the per-city non-null feature availability matrix, computed by scripts/feature_availability.py. The matrix is the design constraint behind the feature-tier registry: cells with 0% availability on a given city are hard blockers and force separate per-country 15
Table 5: T2 canonical property-type scripts/property_type_mapping.json.
taxonomy
(22
categories).
Mapping:
Category
Typical raw forms mapped here
Cities where populated
Office
”Office”, ”Medical Office”, ”Bank Branch”
nyc, la, seattle, dc, chicago, boston, portland all US + singapore
”Multifamily Housing”, ”Multifamily”, ”Residential”, ”Mixed Residential”, ”Other Lodging/Residential” Retail ”Retail Store”, ”Mall”, ”Shopping Center” (NOT bare ”Commercial”) Hotel ”Hotel”, ”Motel”, ”Lodging” K-12 School ”K-12 School”, ”School” College/University ”College/University”, ”University” Hospital/Medical ”Hospital”, ”Medical Office”, ”Clinic” Warehouse/Distribution”Warehouse”, ”Distribution Center” Industrial ”Manufacturing/Industrial”, ”Industrial” Worship ”Worship Facility”, ”Church”, ”Synagogue”, ”Mosque” Senior Living ”Senior Living Community”, ”Nursing Home” Supermarket/Grocery ”Supermarket/Grocery Store” Restaurant ”Restaurant”, ”Food Service” Parking ”Parking”, ”Garage” Laboratory ”Laboratory” Fitness Center ”Fitness Center/Health Club” Library ”Library” Residence Hall ”Residence Hall/Dormitory” ”Mixed Use Property”, Mixed Use ”RES/COMMERCIAL USE” Self-Storage ”Self-Storage Facility” ”Commercial”, ”Commercial - Port Facility” Commercial (Mixed)
Multifamily Housing
Other
”Other”, ”Mixed Use - Commercial”, residuals
nyc, la, seattle, dc, chicago all US nyc, dc, chicago, seattle nyc, dc, chicago all US nyc, la, seattle, dc nyc, la nyc, la, seattle, dc most US la, chicago, seattle small la, dc seattle, chicago small dc nyc, dc nyc, la, dc nyc, la sf only – SF raw data uses one coarse ”Commercial” bucket that covers offices, retail, and mixed-use; we keep it as a separate category to avoid inflating the Retail share. all
tiers (notably AU has 0% property_type and year_built in the NABERS/BEEC schema, and Singapore has 0% site_eui but 100% source_eui).
H
Building Feature-tier Registry
Table 6 lists the feature-tier registry used by the building track. Clean tiers are the deployable settings; proxy-rich and direct-energy-proxy tiers are retained only to quantify the value and risk of richer disclosure fields. Table 7 reports the same tier ladder as a numerical summary, complementing Figure 3 in the main text.
I
Statistical Protocol
Hyperparameter tuning. Tree baselines are tuned with validation-based random search: 40 trials selecting by validation R2 on the strict-coverage panel for the company track, and 15 trials selecting by validation log-MAE on the in-distribution building task. The MLP uses the same 15-trial budget over the search space listed in scripts/run_task_a_mlp_hpsearch.py. Best retained configurations per (feature tier, model) are listed in Appendix J. Multi-seed reporting. Headline company Scope 1+2 baselines are reported with 5-seed mean and standard deviation. The building-track headline regression (Table 2) and main-text feature-tier ladder (Figure 3) are 5-seed mean±std; the appendix per-tier feature ladder (Figure 6) and per-city cross-city transfer tables report 3-seed mean and standard deviation, since the 2-seed extension was applied 16
Figure 5: Per-city non-null availability (%) for building-level schema fields. Cells at 0% are hard exclusions for the corresponding tier.
overall R 2 (3-seed mean ± std)
Task A: feature-tier ladder, grouped-building split 0.8 0.6
RandomForest XGBoost
LightGBM MLP
0.4 0.2 0.0 leakage-prone high-leakage
0.2 0.4
core c
e
mat
cli ore+
ore
us_c
eta
us_m
us_e
ui
us_f
ull
au_c
ore
ui
au_e
ull
au_f
Figure 6: building-track regression feature-tier ladder, grouped-building split, 3-seed mean±std. Hatching: // proxy-rich, xx direct-energy-proxy. MLP init-only 5-seed refit: Table 11.
only to the headline grouped-split panel. The MLP combines split-induced and initialisation-induced variance; an init-only 5-seed refit on a fixed split is reported in Appendix N (Table 11), and single-seed MLP numbers are not used for model-ordering claims. Uncertainty and significance testing. Per-metric uncertainty is reported as a 1000-sample bootstrap confidence interval. Headline model-ordering claims—company structured-vs-LLM, tuned-treefamily internal comparisons, and building in-distribution model-ordering—are accompanied by a 1000-sample paired bootstrap on shared test rows. Full pair-wise ∆R2 values and p-values are tabulated in Appendix S; we treat p < 0.05 as the threshold for explicit ordering claims, and absence of paired-bootstrap evidence is reported as a tie rather than as a model-superiority claim.
J
Hyperparameter Search and Best Configurations
building-track regression tree baselines are tuned with 15-trial validation-based random search; the validation metric is log-MAE on a held-out validation slice created by the same grouped-building 17
Table 6: Building-track feature tiers. Clean tiers exclude energy-use proxies; proxy-rich tiers add EUI/rating; direct-energy-proxy tiers add raw electricity/gas. Proxy tiers are never pooled with clean tiers. Tier
Cities
Added features beyond size, coordinates, year, HDD, and CDD
Paper label
Role in paper
core_all_cities
26
None.
Clean
core_all_cities_climate_plus
26
Clean
us_core
6 U.S.
Annual mean temperature, relative humidity, surface solar radiation, and wind speed from NASA POWER. Property type and year built.
Main cross-country headline tier. Climate ablation against the core tier.
us_metadata
6 U.S.
us_core plus STAR score.
us_leaky_eui
6 U.S.
us_metadata plus site and source EUI.
Proxy-rich
us_leaky_full
6 U.S.
au_core
15 AU
Direct-energy proxy Clean
au_eui
15 AU
us_leaky_eui plus electricity and natural-gas use. Climate-plus core features only. au_core plus site EUI.
au_full
15 AU
au_eui plus a NABERSderived rating score.
Proxy-rich
ENERGY
Clean
Clean
Proxy-rich
U.S. metadata-rich baseline with fixed city population. Tests non-energy metadata value in U.S. disclosures. Upper-reference tier; not a deployable headline result. Upper-reference tier with proximate target signals. Australia-only cleanschema comparison. Australia-only EUI reference tier. Australia-only rating/EUI reference tier.
Table 7: Best point-estimate grouped-building R2 on T2 Task A across the nine feature tiers (best of tuned trees and TabPFN v2 per tier). Tier
Coverage / added signal
Best model
R2
Cross-country core Cross-country + weather U.S. core U.S. metadata U.S. EUI proxy U.S. direct-energy proxy AU core AU EUI proxy AU rating/EUI proxy
26 cities, clean physical core Core + NASA POWER weather 6 U.S. cities, property/year metadata U.S. core + ENERGY STAR score U.S. + EUI proxy fields U.S. + electricity/gas proxies 15 AU metros, clean AU schema AU + EUI proxy field AU + EUI and NABERS-derived rating
TabPFN v2 TabPFN v2 TabPFN v2 TabPFN v2 RandomForest TabPFN v2 RandomForest TabPFN v2 TabPFN v2
0.479 ± 0.024 0.480 ± 0.022 0.511 ± 0.019 0.563 ± 0.032 0.747 ± 0.010 0.821 ± 0.028 0.438 ± 0.050 0.591 ± 0.083 0.615 ± 0.069
split as the test set. Building-track MLP baselines use the same 15-trial budget over the search space listed in scripts/run_task_a_mlp_hpsearch.py. Company-track tree baselines use a 40-trial random search selecting by validation R2 on the strict-coverage panel. Table 8 reports the best configuration retained for each (feature tier, model) on building-track regression, together with the test-set R2 and log-MAE produced by re-fitting that configuration. Differences within the tuned tree family are small, consistent with the in-tier paired-bootstrap results in Section 5.2. Table 8: T2 Task A best HP configs (15-trial random search per (tier, model), selected by validation log-MAE). Feature tier
Model
Trials
Test R2
Test LogMAE
au_core
LightGBM
15
0.421
0.411
au_core
MLP
15
0.392
au_core au_core
RandomForest XGBoost
15 15
0.425 0.410
au_eui
LightGBM
15
0.551
Best configuration
n_estimators=800, num_leaves=31, learning_rate=0.08, min_child_samples=20, reg_alpha=1.0, reg_lambda=0.1, colsample_bytree=0.7, subsample=1.0, subsample_freq=1 0.409 hidden_layers=[512, 512, 256], dropout=0.1, lr=0.003, weight_decay=0.0001, batch_size=2048, max_epochs=200, patience=20 0.399 n_estimators=500, max_depth=10, min_samples_leaf=5, max_features=0.5 0.418 n_estimators=800, max_depth=8, learning_rate=0.05, min_child_weight=3.0, subsample=1.0, colsample_bytree=1.0, reg_alpha=0.0, reg_lambda=5.0 0.209 n_estimators=800, num_leaves=255, learning_rate=0.05, min_child_samples=5, reg_alpha=0.1, reg_lambda=5.0, colsample_bytree=1.0, subsample=0.7, subsample_freq=1
18
Table 8 – continued Feature tier
Model
Trials
Test R2
Test LogMAE
au_eui
MLP
15
0.540
0.213
au_eui au_eui
RandomForest XGBoost
15 15
0.555 0.549
0.217 0.203
au_full
LightGBM
15
0.560
0.203
au_full
MLP
15
0.513
0.228
au_full au_full
RandomForest XGBoost
15 15
0.565 0.555
0.214 0.204
core
LightGBM
15
0.421
0.622
core
MLP
15
-0.421
0.651
core core
RandomForest XGBoost
15 15
0.437 0.410
0.620 0.622
core+climate
LightGBM
15
0.419
0.624
core+climate
MLP
15
-0.453
0.657
core+climate core+climate
RandomForest XGBoost
15 15
0.422 0.412
0.623 0.621
us_core
LightGBM
15
0.433
0.576
us_core
MLP
15
0.139
0.603
us_core us_core
RandomForest XGBoost
15 15
0.519 0.451
0.574 0.570
us_leaky_eui
LightGBM
15
0.763
0.150
us_leaky_eui
MLP
15
0.457
0.213
us_leaky_eui us_leaky_eui
RandomForest XGBoost
15 15
0.789 0.740
0.139 0.142
us_leaky_full
LightGBM
15
0.794
0.097
us_leaky_full
MLP
15
0.210
0.117
us_leaky_full us_leaky_full
RandomForest XGBoost
15 15
0.854 0.825
0.076 0.085
us_metadata
LightGBM
15
0.493
0.470
us_metadata
MLP
15
0.472
0.485
us_metadata us_metadata
RandomForest XGBoost
15 15
0.577 0.464
0.453 0.463
K
Best configuration hidden_layers=[256, 128, 64], dropout=0.0, lr=0.001, weight_decay=0.0001, batch_size=1024, max_epochs=200, patience=20 n_estimators=500, max_depth=20, min_samples_leaf=2, max_features=0.7 n_estimators=800, max_depth=4, learning_rate=0.08, min_child_weight=1.0, subsample=0.8, colsample_bytree=0.9, reg_alpha=1.0, reg_lambda=5.0 n_estimators=800, num_leaves=127, learning_rate=0.08, min_child_samples=20, reg_alpha=1.0, reg_lambda=0.1, colsample_bytree=1.0, subsample=0.8, subsample_freq=1 hidden_layers=[256, 256, 128], dropout=0.1, lr=0.003, weight_decay=1e-05, batch_size=2048, max_epochs=200, patience=20 n_estimators=200, max_depth=20, min_samples_leaf=1, max_features=0.7 n_estimators=800, max_depth=8, learning_rate=0.05, min_child_weight=10.0, subsample=1.0, colsample_bytree=1.0, reg_alpha=1.0, reg_lambda=5.0 n_estimators=800, num_leaves=127, learning_rate=0.05, min_child_samples=5, reg_alpha=1.0, reg_lambda=0.1, colsample_bytree=1.0, subsample=0.9, subsample_freq=1 hidden_layers=[256, 256, 128], dropout=0.1, lr=0.003, weight_decay=0.0001, batch_size=4096, max_epochs=200, patience=20 n_estimators=500, max_depth=20, min_samples_leaf=5, max_features=0.5 n_estimators=800, max_depth=8, learning_rate=0.05, min_child_weight=3.0, subsample=1.0, colsample_bytree=1.0, reg_alpha=0.0, reg_lambda=5.0 n_estimators=800, num_leaves=127, learning_rate=0.05, min_child_samples=20, reg_alpha=0.1, reg_lambda=5.0, colsample_bytree=0.8, subsample=1.0, subsample_freq=1 hidden_layers=[128, 64, 32], dropout=0.0, lr=0.003, weight_decay=0.0001, batch_size=2048, max_epochs=200, patience=20 n_estimators=200, max_depth=15, min_samples_leaf=5, max_features=0.5 n_estimators=800, max_depth=10, learning_rate=0.02, min_child_weight=1.0, subsample=0.7, colsample_bytree=0.7, reg_alpha=1.0, reg_lambda=5.0 n_estimators=800, num_leaves=127, learning_rate=0.08, min_child_samples=20, reg_alpha=1.0, reg_lambda=0.1, colsample_bytree=1.0, subsample=0.8, subsample_freq=1 hidden_layers=[512, 256, 128], dropout=0.1, lr=0.001, weight_decay=0.0, batch_size=4096, max_epochs=200, patience=20 n_estimators=500, max_depth=25, min_samples_leaf=2, max_features=0.5 n_estimators=800, max_depth=10, learning_rate=0.02, min_child_weight=10.0, subsample=0.7, colsample_bytree=0.9, reg_alpha=0.0, reg_lambda=1.0 n_estimators=800, num_leaves=63, learning_rate=0.1, min_child_samples=5, reg_alpha=0.0, reg_lambda=0.0, colsample_bytree=0.7, subsample=1.0, subsample_freq=1 hidden_layers=[512, 512, 256], dropout=0.1, lr=0.001, weight_decay=0.0001, batch_size=4096, max_epochs=200, patience=20 n_estimators=200, max_depth=25, min_samples_leaf=2, max_features=0.5 n_estimators=800, max_depth=10, learning_rate=0.1, min_child_weight=3.0, subsample=1.0, colsample_bytree=1.0, reg_alpha=1.0, reg_lambda=1.0 n_estimators=800, num_leaves=127, learning_rate=0.1, min_child_samples=10, reg_alpha=0.0, reg_lambda=0.0, colsample_bytree=1.0, subsample=1.0, subsample_freq=1 hidden_layers=[512, 256, 128], dropout=0.0, lr=0.0005, weight_decay=0.0, batch_size=1024, max_epochs=200, patience=20 n_estimators=500, max_depth=25, min_samples_leaf=1, max_features=0.7 n_estimators=800, max_depth=10, learning_rate=0.05, min_child_weight=10.0, subsample=0.9, colsample_bytree=0.8, reg_alpha=0.0, reg_lambda=0.1 n_estimators=800, num_leaves=31, learning_rate=0.1, min_child_samples=10, reg_alpha=0.0, reg_lambda=5.0, colsample_bytree=0.9, subsample=0.7, subsample_freq=1 hidden_layers=[256, 128, 64], dropout=0.1, lr=0.001, weight_decay=1e-05, batch_size=2048, max_epochs=200, patience=20 n_estimators=300, max_depth=25, min_samples_leaf=2, max_features=0.5 n_estimators=800, max_depth=10, learning_rate=0.08, min_child_weight=5.0, subsample=0.9, colsample_bytree=0.8, reg_alpha=1.0, reg_lambda=0.1
Building-Track Feature Attribution
Table 9 compares permutation importance and SHAP attribution for the tuned LightGBM model on the 26-city core_all_cities tier. Both diagnostics identify gross floor area as the dominant signal, followed by location; year and climate-derived covariates are smaller on this tuned tree model, matching the main-text finding that explicit climate fields change tuned-tree R2 only marginally on the core tier.
L
Company-Track Feature Attribution
Figure 7 and Table 10 report the attribution diagnostic for the tuned LightGBM model on the company track (the strict-coverage panel). Revenue dominates with permutation ∆R2 = 0.652, more than twice the next feature; the ExioML sectoral factor is second at 0.266, and EBITDA, employees, and market cap form a similar-magnitude secondary group. The only sector dummy that enters the top six is GICS Financial Services, consistent with the within-GICS sub-sector heterogeneity discussed in the sector-factor case study (Appendix O.1). Mean |SHAP| rankings agree with permutation order. 19
Table 9: T2 Task A feature attribution on core_all_cities, tuned LightGBM. Permutation: R2 drop after shuffling (5 repeats). SHAP: mean |SHAP| on 5,000 test rows. Feature
Permutation ∆R2
Mean |SHAP|
Gross floor area Longitude Latitude Heating degree days Reporting year Cooling degree days
0.678 ± 0.014 0.167 ± 0.019 0.154 ± 0.013 0.042 ± 0.003 0.033 ± 0.005 0.025 ± 0.003
0.729 0.346 0.177 0.045 0.095 0.033
5 H Y H Q X H
( [ L R 0 / I D F W R U
( % , 7 ' $
( P S O R \ H H V
0 D U N H W F D S
* , & 6 ) L Q a 6 Y F
2 3 H U P X W D W L R Q Rlog
Figure 7: Tuned LightGBM permutation ∆R2 on the strict-coverage panel (top six features; error bars: std over five repeats).
M
Compute and Runtime
All experiments were run on a single workstation with 8 × NVIDIA RTX A5000 (24 GB) GPUs and a multi-core CPU; only TabPFN, MLP, and time-series foundation-model inference made use of GPUs. Tree baselines (RandomForest, XGBoost, LightGBM, HistGradientBoosting) ran exclusively on CPU. The dominant cost is the multi-seed re-fit campaigns rather than hyperparameter search. Indicative wall-clock numbers from the experiment logs: • building-track regression 3-seed re-fit across the nine feature tiers and four model families: about 4–6 hours per seed on CPU plus ≤ 1 GPU-hour per seed for MLP. • Building-track regression + Sentinel-2 3-seed re-fit on the S2-eligible subset: about 2–3 hours per seed because the S2-aligned subset is smaller. • Building-track cross-city transfer 3-seed re-fit (23 held-out cities, four model families): about 12–18 hours per seed on CPU. • Company-track tuned tree family (40-trial search per model on the strict-coverage panel): ≤ 1 CPUhour per model. • Company-track LLM panel (200-row stratified sample, four LLMs × three prompt regimes): a few API-hours total at standard rate limits. A full reproduction of every multi-seed table reported in the main text takes on the order of 80–120 CPU-core-hours plus a few GPU-hours, well within a single-workstation budget. Per-experiment elapsed times are stored in elapsed_sec columns of the per-run CSVs and in the script logs.
N
MLP Initialisation-only Multi-seed Refit
Table 11 reports a fixed-split init-only refit (train/validation/test fold held constant; five seeds vary only the network initialisation). This complements the headline 5-seed protocol used in §5.2 and Figure 3, where each seed draws a fresh building-grouped split. Single-seed MLP/FT-Transformer R2 on small panels reflects optimisation variance; the per-seed table is the more informative artefact for downstream users. 20
Table 10: T1-A Scope 1+2 feature attribution on T1-Strict, tuned LightGBM. Permutation: R2log drop after shuffling (5 repeats). SHAP: mean |SHAP| on full test panel. Top 6 by permutation rank. Feature Revenue ExioML factor EBITDA Employees Market cap GICS Financial Services
Permutation ∆R2log
Mean |SHAP|
0.652 ± 0.016 0.266 ± 0.013 0.207 ± 0.008 0.178 ± 0.011 0.161 ± 0.007 0.035 ± 0.003
0.477 0.298 0.203 0.124 0.110 0.070
Table 11: MLP init-only refit on T2 Task A grouped split: 5 seeds vary network init only (split fixed). On core_all_cities, 4/5 seeds yield negative R2 . Feature tier
nseeds
R2 mean
R2 std
R2 min
R2 max
MAE mean
5 5 5 5 5 5
-0.294 -0.255 0.266 0.415 0.552 0.228
0.332 0.337 0.208 0.154 0.175 0.231
-0.502 -0.500 0.033 0.187 0.379 -0.051
0.290 0.340 0.498 0.543 0.765 0.478
415.1 408.3 361.1 302.7 184.9 151.3
core_all_cities core_all_cities_climate_plus us_core us_metadata us_leaky_eui us_leaky_full
O
Company-Track Supplement
Table 12 reports the full Scope 1+2 structured baseline panel summarised by Table 1. Table 13 reports the LLM-only panel summarised in Section 5.1. Table 14 reports the Scope 3 regression panel and the one-step annual forecasting panel summarised in Section 5.1; the company panel spans only six reporting years (2018–2023), which is too short to support a meaningful multi-horizon protocol after train/val/test allocation, so company-track forecasting is reported at a single horizon only. Multi-horizon evaluation is provided on the longer building panel (Table 16). HistGradientBoosting hyperparameters on Scope 3 (the matched-panel Scope 3 task) are tuned independently with a 40-trial random search on the filtered Scope 3 panel rather than reusing the Scope 1+2 configuration; the retuned model reaches R2 = 0.653 in Table 1. O.1
Sector-Factor Case Study
The sector-factor baseline in Table 1 captures coarse scale but can miss firm-level magnitude by orders of magnitude. A representative U.K. financial company in the panel reports $8.5 B revenue. The ExioML factor for Financial Services is 10.7 tCO2 e per million USD, so the factor baseline predicts 90,841 tCO2 e. The actual reported Scope 1+2 is 750 tCO2 e, a 121× over-prediction. Banks, insurers, asset managers, and payment firms all fall under broad Financial Services labels, but their operating-emission intensities differ substantially; a single sectoral factor therefore cannot resolve lower-intensity sub-sectors. Finer-grained sector labels partially reduce this error but do not close the gap. Recomputing the intensity factor on training rows at three SICS granularities for the same company—a private-equity asset manager—yields 8.98, 7.69, and 5.15 tCO2 e per million USD at the GICS-11, SICS sector, and SICS sub-sector (Capital Markets) levels, reducing the over-prediction from 121× to 58×. Going one level finer to SICS industry (Asset Management & Custody Activities, 92 training rows) reverses the trend to 111×, because the partition is too small to estimate a stable factor. On the full 152-row U.K. Financial Services cohort, geometric-mean over-prediction drops from 6.9× at GICS-11 to 4.6× at sub-sector and rises back to 7.5× at industry. Consistent with this pattern, adding sub-sector target encoding to the strict-coverage panel raises default LightGBM by ∆R2 = +0.021, while extending the encoding to industry adds essentially no gain (−0.0004).
P
Building-Track Transfer and Forecasting Summary
Table 15 collects the building-track stress-test summaries: strict temporal hold-out, City-LOCO transfer, the Sentinel-2 cross-city delta, and one-step forecasting. We keep these in the appendix 21
Table 12: T1 Scope 1+2 headline structured baselines on T1-Strict. LLM panel: Table 13. Subset
Model
ntest
MAElog
R2log
Pearson r
T1-Wide
Sector-factor lookup LightGBM (default) XGBoost (default)
1329 3135 3135
0.878 0.933 0.933
0.222 0.300 0.303
0.683 0.548 0.551
T1-Strict
Sector-factor lookup LightGBM (default) XGBoost (default) TabPFN v2 MLP (neural) FT-Transformer (neural) LightGBM (tuned) (5-seed σR2 =0.0013) XGBoost (tuned) RandomForest (tuned) HistGradientBoosting (tuned)
1329 1363 1363 1363 1363 1363 1363 1363 1363 1363
0.878 0.340 0.459 0.262 0.498 0.565 0.219 0.207 0.226 0.243
0.222 0.826 0.732 0.857 0.693 0.622 0.873 0.879 0.875 0.871
0.683 0.911 0.859 0.927 0.834 0.790 0.935 0.938 0.936 0.933
Table 13: T1 Scope 1+2 LLM panel. Best prompt regime per model by log-MAE; paired ∆R2 = RandomForest tuned − LLM on overlapping subset. LLM
Prompt
n
MAElog
R2log
∆R2 vs RF
p
Claude Haiku 4.5 GPT-4o mini Qwen2.5-7B Instruct Mistral-7B Instruct
zero-shot zero-shot few-shot few-shot
200 200 200 200
0.616 0.738 0.930 1.159
0.586 0.491 0.213 -0.201
0.323 0.422 0.702 1.121
< 0.001 < 0.001 < 0.001 < 0.001
because they diagnose failure modes and regime shifts rather than define the strongest headline in-distribution baselines. Temporal hold-out is comparatively stable. Task B1 tree models retain overall R2 ≈ 0.40–0.50 on core_all_cities, broadly comparable to their grouped-building Task A scores; strict year hold-out is much easier than the cross-city transfer of Section 5.2. Forecasting has several competitive families. On the forecasting panel, TimesFM reaches the highest displayed R2 (0.842), Chronos has the lowest MAE among the displayed models (122.1), RandomForest has the lowest log-MAE (0.228), and tree-family lag models range from 0.705 to 0.754 in R2 . The benchmark thus surfaces three distinct axes: cross-city transfer is difficult, strict year hold-out is comparatively stable, and forecasting is contested across baseline families rather than dominated by a single class. Multi-horizon forecasting. The Task E protocol fixes train end at year 2019 but evaluates every test year y ≥ 2021 separately, so the same trained model is queried at horizons h = y−2019 ranging from h = 2 (2021) to h = 5 (2024). Per-year R2 is in Table 16. Time-series foundation models (TimesFM, Chronos) hold up best at the longer horizons, reaching R2 =0.90 at h=4; lag-feature gradient-boosted regressors plateau slightly lower (0.66–0.81); MLP degrades fastest (0.54 at h = 2 to 0.39 at h = 5). The horizon decay is mild because the lag-feature input remains the immediately preceding year (i.e. each prediction is one-step from its own past), so what is varied is mainly distributional drift between train end and test year rather than recursive multi-step error accumulation.
Q
Task B Full Model Grid
The main text reports Task B as a brief observation: strict year hold-out does not collapse on core_all_cities, with tree models retaining overall R2 in the 0.40–0.50 range. Table 17 provides the complete model grid for both the B1 strict-year hold-out and the B2 COVID variant, including the auxiliary baselines that the main text omits. The B2 COVID variant trains on years ≤ 2018 and evaluates 2019, 2020, and 2021 separately, isolating pandemic-year disruptions from in-distribution temporal drift. 22
Table 14: T1 Scope 3 prediction (tuned trees on T1-A′ ) and T1-E one-step Scope 1+2 forecasting (balanced 2018–2022 panel). Block
Baseline
ntest
MAElog
R2log
Pearson r
T1-A′ Scope 3
LightGBM (tuned, log target) XGBoost (tuned, log target) RandomForest (tuned, log target)
982 982 982
0.607 0.583 0.594
0.637 0.645 0.639
0.805 0.806 0.803
T1-E forecasting
Persistence Sector growth Three-year moving average Linear trend XGBoost (lag features) Chronos-Bolt small TimesFM 2.0 500M Moirai 1.1 small
1930 1930 1930 1930 386 1930 1930 1930
0.088 0.089 0.119 0.131 0.123 0.147 0.143 0.134
0.959 0.959 0.947 0.936 0.933 0.776 0.808 0.930
0.979 0.979 0.973 0.969 0.966 0.906 0.916 0.965
Table 15: Building-track transfer and forecasting on core_all_cities. Left: temporal hold-out (R2 ) and City-LOCO (mean±std across held-out metros×seeds); +S2 is the PCA-64 Sentinel-2+Clay delta on the matched subset. Right: one-step forecasting with lagged-only inputs. One-step forecasting
Temporal and cross-city Model
Temporal
GlobalMean −0.021 CityTypeMean 0.034 TabPFN v2 0.466 ± 0.014 RF 0.498 LightGBM 0.398 XGBoost 0.389 MLP 0.368
R
City-LOCO mean + S2 (∆)
Model
MAE LogMAE
−0.224 ± 0.20 −0.224 ± 0.20 −0.127 ± 1.07 0.127 ± 0.38 0.030 ± 0.66 0.084 ± 0.60 −2.55 ± 9.83
GlobalMean CityTypeMean LinearTrend RF LightGBM XGBoost MLP Chronos TimesFM Moirai
543.7 495.1 172.4 131.7 143.7 144.9 322.4 122.1 125.5 139.1
— — — — +0.124 +0.053 −0.229
1.164 0.894 0.397 0.228 0.233 0.231 0.822 0.235 0.236 0.234
R2 −0.021 0.048 0.616 0.754 0.715 0.705 0.555 0.819 0.842 0.783
Task D Cross-Property-Type Stress Test
Per Section 5.2, Task D is reported as appendix-level evidence. Table 18 lists the leave-one-propertytype-out performance on the seven main types (Office, Multifamily Housing, Retail, Hotel, K12 School, Hospital/Medical, Warehouse/Distribution) for the us_core and us_metadata tiers, which are the two tiers with reliable property-type coverage across U.S. cities. Performance varies substantially by held-out type, and the metadata tier does not uniformly help every type. Table 18: Task D cross-property-type stress test on us_core and us_metadata: hold out one type at training, evaluate on it. Tier
Held-out type
Model
us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core
Office Office Office Office Office Office Office Office Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Retail
CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean
23
ntest
R2
LogMAE
45650 45650 45650 45650 45650 45650 45650 45650 256168 256168 256168 256168 256168 256168 256168 256168 8152
-1.202 -0.148 0.245 0.033 0.392 -1.132 0.298 0.254 -1.755 -0.048 0.230 0.069 0.340 -6.400 0.296 0.218 -0.915
1.606 1.560 0.939 1.119 1.017 1.409 0.851 0.943 0.996 1.011 0.799 1.059 0.868 0.953 0.739 0.794 1.358
Table 18 – continued Tier
Held-out type
Model
us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_core us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata
Retail Retail Retail Retail Retail Retail Retail Hotel Hotel Hotel Hotel Hotel Hotel Hotel Hotel K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Office Office Office Office Office Office Office Office Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Multifamily Housing Retail Retail Retail Retail Retail Retail Retail Retail Hotel Hotel
GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean
24
ntest
R2
LogMAE
8152 8152 8152 8152 8152 8152 8152 9959 9959 9959 9959 9959 9959 9959 9959 24657 24657 24657 24657 24657 24657 24657 24657 5282 5282 5282 5282 5282 5282 5282 5282 13497 13497 13497 13497 13497 13497 13497 13497 45650 45650 45650 45650 45650 45650 45650 45650 256168 256168 256168 256168 256168 256168 256168 256168 8152 8152 8152 8152 8152 8152 8152 8152 9959 9959
-0.068 0.269 0.119 0.316 0.107 0.350 0.186 -1.636 -0.136 0.183 0.047 0.224 -0.503 0.309 0.164 -7.703 -0.091 0.096 0.040 0.146 -0.075 -0.011 0.092 -0.168 -0.166 0.002 -0.073 0.189 -0.219 0.153 0.002 -13.410 -0.003 0.169 0.154 0.340 0.135 -0.673 0.182 -1.202 -0.148 0.291 0.036 0.386 -0.983 0.364 0.319 -1.755 -0.048 0.398 0.161 0.447 -7.240 0.415 0.401 -0.915 -0.068 0.254 0.181 0.274 0.107 0.373 0.260 -1.636 -0.136
1.402 1.084 1.111 1.169 1.314 1.031 1.098 1.528 1.300 1.010 1.180 1.144 1.364 0.624 1.016 1.126 0.971 0.696 0.787 0.746 1.051 0.773 0.689 2.098 2.021 1.685 1.869 1.663 1.915 1.227 1.681 1.639 1.705 0.973 1.013 1.030 1.150 1.767 0.975 1.606 1.560 1.009 1.183 1.092 1.498 0.734 1.010 0.996 1.011 0.885 1.158 0.962 0.910 0.583 0.884 1.358 1.402 1.247 1.164 1.327 1.293 0.993 1.236 1.528 1.300
Table 18 – continued
S
Tier
Held-out type
Model
us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata us_metadata
Hotel Hotel Hotel Hotel Hotel Hotel K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School K-12 School Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Hospital/Medical Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution Warehouse/Distribution
LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge TabPFN v2 XGBoost
ntest
R2
LogMAE
9959 9959 9959 9959 9959 9959 24657 24657 24657 24657 24657 24657 24657 24657 5282 5282 5282 5282 5282 5282 5282 5282 13497 13497 13497 13497 13497 13497 13497 13497
0.320 0.129 0.303 -0.463 0.348 0.307 -7.703 -0.091 0.209 0.144 0.218 -0.571 -0.114 0.200 -0.168 -0.166 0.158 -0.040 0.241 -0.147 0.173 0.147 -13.410 -0.003 0.461 0.329 0.357 0.067 -0.803 0.450
0.987 1.206 1.028 1.288 0.605 0.984 1.126 0.971 0.703 0.735 0.727 1.058 0.894 0.709 2.098 2.021 1.766 1.942 1.815 1.896 1.151 1.768 1.639 1.705 0.754 0.737 0.972 1.062 1.805 0.749
Paired-bootstrap and Per-city Tables
Full paired-bootstrap deltas and p-values for every (feature tier, model pair) on building-track regression are released as the canonical CSVs results/clean_building/task_a_pairwise_bootstrap.csv (US and crosscountry, tree-vs-tree), task_a_pairwise_bootstrap_au.csv (AU, tree-vs-tree), and task_a_pairwise_bootstrap_with_tabpfn.csv (TabPFN v2 vs each tuned tree, all six tiers; Table 20). Table 20 summarises the TabPFN-vs-tree comparison that backs the main-text claim in §5.2. Figure 10 visualises the ∆R2 matrix together with significance stars; the CSVs additionally expose paired deltas on MAE and log-MAE, the win-rate fractions on bootstrap draws, and the bootstrap ntest used per cell. RandomForest’s grouped-split advantage is schema-dependent, not universal. On the four U.S.-only tiers from clean metadata through target-proxy fields, RandomForest exceeds LightGBM by 0.026–0.086 R2 with p ≤ 0.018, and exceeds XGBoost on three of those four tiers at p < 0.05. On the two cross-country tiers the effect is smaller and less consistent, and on all three AU-only tiers every tree-vs-tree pair has pR2 ≥ 0.088. The natural reading is that RandomForest’s bagging advantage is strongest inside the heterogeneous U.S. metadata ladder, weakens under deliberately minimal cross-country schemas, and disappears on Australian disclosures, which share a single NABERS/BEEC schema. Per-city stratified R2 for building-track regression is released as task_a_stratified_r2.csv (USonly) and task_a_stratified_r2_au.csv (AU-only); each row is a (feature tier, model, city) triple with three-seed mean and standard deviation. The 522 rows are too long for a paper table, so they remain in the released CSVs. Table 19 reports the per-held-out-city R2 values behind the main-text City-LOCO summary on the core_all_cities tier (Figure 4). The table keeps the raw pathological values for all models including tabular MLP, which is omitted from the main figure. 25
Table 16: T2 Task E multi-horizon forecasting on core_all_cities: train on years ≤2019, validate on 2020, evaluate each test year y at horizon h = y − 2019. Each cell is per-year R2 . Trees and TabPFN-style baselines use lagged emissions and lagged climate; foundation models (Chronos / TimesFM / Moirai) consume the lagged emissions sequence directly. Ridge collapses on this lagfeature setup at every horizon (heavy negative; omitted from comparison). h = 2 (2021)
h = 3 (2022)
h = 4 (2023)
h = 5 (2024)
−0.02 +0.06 +0.61
−0.02 +0.05 +0.67
−0.03 +0.05 +0.47
−0.02 +0.02 +0.73
Time-series foundation models Chronos +0.73 TimesFM +0.77 Moirai +0.74
+0.84 +0.88 +0.72
+0.90 +0.90 +0.86
+0.81 +0.82 +0.82
Lag-feature regressors RandomForest +0.68 LightGBM +0.67 XGBoost +0.66 MLP +0.54
+0.78 +0.74 +0.72 +0.66
+0.81 +0.77 +0.78 +0.59
+0.74 +0.67 +0.64 +0.39
Model Naive references GlobalMean CityTypeMean LinearTrend
Table 17: Task B on core_all_cities. B1: train ≤ 2019, val 2020, test ≥ 2021. B2: train ≤ 2018, evaluate 2019/2020/2021 separately.
T
Model
B1 ntrain
B1 ntest
B1 R2
B2 R22019
B2 R22020
B2 R22021
CityTypeMean GlobalMean LightGBM MLP RandomForest Ridge XGBoost
212003 212003 212003 212003 212003 212003 212003
212279 212279 212279 212279 212279 212279 212279
0.034 -0.021 0.398 0.368 0.498 -1.322 0.389
0.061 -0.023 0.512 0.426 0.596 -1.521 0.502
0.072 -0.020 0.488 0.371 0.568 -2.301 0.472
0.050 -0.015 0.416 0.256 0.495 -1.582 0.400
Sentinel-2 + Clay Supplementary Variants
The headline Sentinel-2 figure (Figure 9) excludes the PCA-128 and gated two-tower fusion variants because they are either single-seed exploratory or do not have multi-seed coverage on both Task A and Task C1. Single-seed numbers for these variants are released as task_a_results_s2_all.csv and task_c_results_s2_all.csv; per-city S2 gain deltas under LOCO are released as task_c1_s2_3seeds_city_deltas.csv so that readers interested in the heterogeneous geographic effect of Sentinel-2 can inspect the per-target-city deltas directly.
U
LLM Panel: Prompt Templates and Failure Modes
The Company-track LLM panel uses one shared prompt scaffold across three regimes (zero-shot, fewshot, no-summary), constructed by scripts/run_t1_a_llm.py. The scaffold has four blocks: a system instruction, an optional labelled-example block (few-shot only), a target company description, and a closing line that invites a single numeric answer. System instruction (verbatim). You are an expert in corporate greenhouse-gas accounting. Given the company information below, predict Scope 1 + Scope 2 emissions in metric tons of CO2-equivalent (tCO2e) for the reporting year. Respond with a SINGLE NUMBER ONLY –- the predicted tCO2e value. Do not include units, explanations, or punctuation. Example: 42500
Company description block. Each company line set covers Company, Country (ISO-2), Sector (GICS-11), Reporting year, Annual revenue (in millions of USD with FX correction), and an optional 26
0.75 0.50 0.25 0.00 0.25 0.50
-1.32
t t ean ean Ridge ores Boos tGBM F XG igh alM ypeM m b o o L T d Gl City Ran
P
ML
1.0
Task E: short-horizon forecasting baselines
0.8 0.6 0.4 0.2 0.0 -8.87
0.2
last v al linea ue r tren d globa l city-t mean ype m ean ridge RF XGB LGBM MLP Chro nos Time sFM Moir ai
Task E1 overall R 2 (one-step forecast; clipped at 0.2)
2021 test) 2019 train, Task B1 overall R 2 (years
Task B: strict temporal hold-out, core\_all\_cities
1.00
Figure 8: Left: Task B1 strict temporal hold-out R2 on core_all_cities (single run). Right: Task E1 short-horizon forecasting R2 . Both panels clipped on the negative side; raw Ridge values annotated. Table 19: Per-held-out-city R2 for T2 Task C1 City-LOCO on core_all_cities, 3-seed mean. Held-out city
RF
LightGBM
Adelaide 0.089 Boston 0.169 Brisbane 0.376 Cairns 0.518 Canberra 0.192 Chicago 0.216 Darwin 0.249 DC 0.230 Denver 0.275 Gold Coast 0.233 Hobart −0.601 Los Angeles 0.120 Melbourne 0.088 Newcastle 0.548 New York City 0.279 Perth 0.406 Philadelphia −0.099 Portland 0.340 Seattle −1.227 San Francisco 0.450 Singapore −0.087 Sydney 0.187 Townsville 0.303 Wollongong −0.213
0.051 0.009 0.485 0.677 0.149 0.257 −0.697 0.207 0.371 0.262 −2.511 0.019 0.040 0.534 0.210 0.436 0.086 0.164 −0.972 0.343 −0.017 0.193 0.322 0.111
XGBoost
MLP
−0.009 0.153 0.017 0.398 0.440 0.486 0.693 0.462 0.177 −0.130 0.452 0.407 −0.202 −0.623 0.198 0.276 0.375 0.305 0.277 0.269 −2.511 0.258 −0.100 −0.085 0.052 −0.088 0.541 0.483 0.204 0.272 0.423 −36.303 0.076 0.097 0.143 0.173 −0.131 −0.910 0.187 0.186 −0.011 −0.163 0.210 −27.290 0.301 0.275 0.213 −0.023
800-character Business description. Few-shot mode prepends ten labelled examples drawn one per GICS-11 sector from a held-out fewshot pool, each followed by the literal line Scope 1+2 emissions: {value} tCO2e. The no-summary regime drops the business-description line. Parsing and failure modes. The response is parsed by extracting the first numeric token (regular expression (-?[\d,]+\.?\d*) with optional scientific notation), stripping commas, and casting to a float. Predictions that fail to parse are flagged NaN and excluded from the test-set metric computation. The most common failure modes observed during runs are: (i) the LLM emits a sentence rather than a number when the description is unusually short; (ii) the LLM returns a range (e.g. 40000–50000); (iii) the LLM returns scientific notation without an exponent (e.g. 4.5e). The parser tolerates (iii) and recovers the numeric prefix, but (i) and (ii) yield NaN; the fraction of NaN responses per (model, regime) is recorded in the per-prediction CSVs released with the benchmark.
V
Reproducibility and Release
The project repository contains the executable reconstruction and evaluation code, canonical split definitions, feature-tier registries, harmonisation metadata, and scripts for regenerating the paper 27
Table 20: T2 Task A paired bootstrap (1,000 resamples on shared grouped-building test rows, seed 42): TabPFN v2 minus each tuned tree. Positive ∆R2 favours TabPFN. p-values are two-sided. Bold entries are significant at p < .05. ∆R2 (TabPFN − tree) [95% CI], p Tier
ntest
1.0
Task A: pooled grouped, S2-eligible subset
0.8 0.6 0.4 0.2 0.0 0.2 0.4
tab MLP core
vs XGBoost
vs RandomForest
94,875 +.025 [+.005, +.044], .012 +.036 [+.014, +.058], .002 +.010 [−.011, +.028], .362 94,875 +.033 [+.018, +.046], .000 +.039 [+.021, +.056], .000 +.029 [+.015, +.043], .000 86,330 +.109 [+.075, +.143], .000 +.090 [+.061, +.121], .000 +.022 [+.001, +.044], .040 86,330 +.108 [+.079, +.140], .000 +.137 [+.088, +.195], .000 +.024 [+.005, +.044], .016 3,704 −.017 [−.108, +.041], .826 −.006 [−.096, +.053], .998 −.020 [−.110, +.037], .740 3,704 +.005 [−.077, +.054], .762 +.007 [−.075, +.058], .744 +.001 [−.084, +.050], .852
LightGBM core
+S2 raw
+S2 PCA64
MLP LightGBM core+climate core+climate
Task C1 mean-city R 2 (city/seed mean ± std)
Task A overall R 2 (3-seed mean ± std)
core (26 cities) core+climate (26 cities) us_core (6 cities) us_metadata (6 cities) au_core (13 cities) au_eui (13 cities)
vs LightGBM
0.5
Task C1: cross-city LOCO, S2-eligible subset
0.0 0.5 1.0 1.5 2.0 2.5 tab
3.0 MLP core
LightGBM core
+S2 raw
+S2 PCA64
MLP LightGBM core+climate core+climate
Figure 9: Sentinel-2 + Clay multimodal extension. Left: Task A grouped R2 (3-seed mean±std) for tab, tab+raw, and tab+PCA-64. Right: Task C1 mean-city R2 across held-out cities and seeds for the same variants. figures and tables from the canonical CSVs. Source data are retrieved from their original portals under the licenses listed in Table 21. Reproducing the full benchmark therefore involves: (i) running the data acquisition step for each portal under its own access terms, (ii) running the canonical preprocessing and feature-tier construction, (iii) running the experiment scripts under scripts/ for each track and task, and (iv) regenerating the figures and tables with scripts/plots/, which read only from the canonical CSVs and emit deterministic outputs.
W
Source Access and License Notes
Table 21 records the source/access plan used for benchmark reconstruction. Raw records are retrieved from their original portals and APIs rather than redistributed wholesale.
28
Task A paired bootstrap: tree pairs × feature tier 0.10 -0.016 -0.003 -0.086***-0.083*** -0.026* -0.060*** -0.004 -0.004 -0.006
0.05
LightGBM-XGBoost +0.011** +0.006 -0.019* +0.029 +0.023* -0.031*** +0.011 +0.002 +0.005 XGBoost-RandomForest -0.027** -0.009 -0.068***-0.112***-0.050***-0.029*** -0.014
0.00
R 2 (modela - modelb)
LightGBM-RandomForest
0.05 -0.005 -0.010
0.10
i i l l core climate us_core s_meta us_eu us_ful au_core au_eu au_ful u + core Figure 10: Paired-bootstrap ∆R2 between tree-family pairs per feature tier. Stars: pR2 < 0.05 (∗), < 0.01 (∗∗), < 0.001 (∗∗∗).
Table 21: Source access and license notes. Raw records are retrieved from the original sources under the listed terms; we release code, harmonisation metadata, and reconstruction recipes. Source
Coverage in GHGbench
Years used
Terms / license checked
Access method
NYC LL84 benchmarking Chicago benchmarking
New York City buildings
2012–2024
NYC Open Data portal / SODA API
Chicago buildings
2014–2023
Seattle benchmarking
Seattle buildings
2015–2024
Washington DC benchmarking San Francisco benchmarking Los Angeles EBEWE Boston BERDO
District of Columbia buildings San Francisco buildings
2013–2024
Los Angeles buildings Boston buildings
2016–2025 2016–2024
Denver benchmarking
Denver buildings
2025
Philadelphia benchmarking Portland benchmarking
Philadelphia buildings
2023
Portland buildings
2019, 2024
Australian Commercial Building Disclosure Singapore BCA building energy performance NASA POWER
15 Australian metropolitan areas Singapore buildings
2011–2026
Building-year meteorology and solar covariates for T2
2011–2025
NYC Open Data / NYC.gov terms of use City data portal open-data terms; dataset metadata did not expose a more specific license in the local archive City of Seattle open-data program and website terms District Data terms; District logos excluded from CC0-style designation DataSF public-domain dedication / PDDL policy DataLA open-data portal terms of use Analyze Boston public-domain dedication / PDDL default Denver Open Data Catalog CC BY 3.0 terms City of Philadelphia website / opendata terms Portland open-data terms; PDDL used for open-data publications where specified Creative Commons Attribution 3.0 Australia on data.gov.au CBD dataset Singapore Open Data Licence via data.gov.sg / BCA publication NASA science data terms; NASA-led data are CC0 unless marked otherwise
CDU
Company Scope 1+2 and Scope 3 disclosures for T1
2018–2023
ExioML / EXIOBASE
Sectoral emission-factor lookup for T1
Sentinel-2 + Clay
T2 satellite embeddings
1995–2022 source panel; matched to T1 reporting years Sentinel-2 L2Aera image composites aligned to S2-eligible building rows
2011–2025
2023–
2017–2020
CDU Terms of Service; noncommercial and download/session restrictions apply EXIOBASE 3 licence information is version-specific on Zenodo; current EXIOBASE3 releases use noncommercial academic terms Copernicus Sentinel free/full/open access terms; Clay model Apache-2.0
29
Chicago Data Portal CSV export
Seattle Open Data portal CSV export Open Data DC CSV export DataSF CSV export Los Angeles Open Data CSV export Analyze Boston CSV/XLSX exports Denver open-data CSV export Philadelphia open-data CSV export Portland open-data XLSX exports
data.gov.au / CBD downloadable dataset BCA / data.gov.sg CSV/XLSX exports NASA POWER daily point API
Reconstruction recipe against CDU exports/API Local ExioML factor-accounting files and lookup script
Sentinel-2 retrieval + Clay embedding extraction pipeline
NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: Abstract and §1 state three benchmark-level findings that match §5: (i) crossentity difficulty asymmetry between company and building tracks; (ii) the in-distribution to out-of-distribution gap dominates within-model variation, with TabPFN v2 paired-bootstrapsignificantly beating every tuned tree on 3/6 building-grouped tiers (Table 20); (iii) multimodal Sentinel-2+Clay embeddings help cross-city transfer but not in-distribution prediction (Figure 4, §5.2). Guidelines: • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: Limitations are discussed in Appendix B, including data-coverage skew toward cities with disclosure laws, address-level geocoding for LA and Denver, the short five-year forecasting panel, and the late-concatenation S2 fusion design. Guidelines: • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate “Limitations” section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 30
3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A] Justification: The paper introduces a benchmark and reports empirical results; it contains no theoretical results requiring proofs. Guidelines: • The answer [N/A] means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and crossreferenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. • Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: All splits, hyperparameter search protocols, multi-seed configurations, and tuned best configs are released in Appendix I and Appendix J, alongside the code repository and Croissant-described dataset on Zenodo (DOI 10.5281/zenodo.20006582). Guidelines: • The answer [N/A] means that the paper does not include experiments. • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). 31
(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Code is released at https://anonymous.4open.science/r/ana_ review-7D14; the building track is published as a Zenodo dataset (DOI 10.5281/zenodo.20006582) under CC-BY-4.0 with a Croissant 1.0 + RAI metadata file; the company track ships as a one-command reconstruction script (§3.1) against a free public API. A self-contained quickstart kit reproduces the headline TabPFN v2 result with one command. Guidelines: • The answer [N/A] means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://neurips.cc/ public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //neurips.cc/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: Hyperparameter search ranges and selected configs are listed in Appendix J; multi-seed protocol and paired-bootstrap construction are in Appendix I; split definitions (stratification, ratios, year cuts) are in Appendix E. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] 32
Justification: Headline numbers are 5-seed mean±std on both tracks; model-ordering claims use 1000-resample paired bootstrap on shared test rows reporting ∆R2 , 95% CI, and twosided p-values. The full statistical protocol is in §4.3 and Appendix I; per-pair tables in Appendix S. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: Appendix M reports compute (RTX A5000 24GB GPUs and multi-core CPU), per-experiment runtimes, and total project compute, including disclosure of preliminary runs that did not appear in the paper. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: The work uses only publicly disclosed building energy/GHG data and free-tier company emissions disclosures, contains no personally identifiable information, and follows the NeurIPS Code of Ethics. Guidelines: • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. 33
• If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: Positive impact: enabling open, comparable evaluation of carbon-emission prediction supports climate-policy and operations research that depended on paywalled disclosures. Negative impact: model errors could be misused to greenwash poor performers; we discuss this risk and the limits of self-reported emissions in Appendix B. Guidelines: • The answer [N/A] means that there is no societal impact of the work performed. • If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: The released assets are tabular emission-disclosure records and Sentinel-2 surface-reflectance embeddings; they pose no high misuse risk (no generative models, no facial recognition, no scraped private content). Guidelines: • The answer [N/A] means that the paper poses no such risks. • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 34
12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: All upstream data sources are cited at first use (§3, Appendix W); each source’s license and terms are noted in Appendix W and respected (open municipal portals are redistributed; CDU is reconstructed not redistributed). Guidelines: • The answer [N/A] means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., CC-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The released building-track dataset ships with a Croissant 1.0 + MLCommons RAI metadata file (14 RAI fields including dataCollection, dataBiases, dataLimitations, dataSocialImpact, hasSyntheticData, and dataReleaseMaintenancePlan), a datasheet (Appendix A), per-file MD5 checksums, and a quickstart kit with expected-output transcript. Guidelines: • The answer [N/A] means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A] Justification: The paper does not involve crowdsourcing or research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. 35
• Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: The paper does not involve crowdsourcing or research with human subjects, so IRB approval is not applicable. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [Yes] Justification: LLMs were used only for writing assistance on the manuscript (typo correction and prose polishing). They did not produce experimental results, design analyses, or generate text for the technical contributions. Guidelines: • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.
36