2025 Machine Learning model update

This year’s Machine Learning model updates show improvements across the board, as a result of increased training and verification data across multiple years.

Expected date of changes: New ML models are already in place, and this will be visible in Emmi universe-data in early Dec 2025

Summary of changes:

• New Machine Learning models trained and validated for public and private equities
• Material Industry Intensity Factors for Scope 3 have been updated with new reported emissions data

New ML Model statistics:

Public Equities machine learning models were retrained with a full 2024 emissions data set, producing better error-rates across the board. These new models will be used for emissions modelling into the future, and as part of this update, were used to re-state 2024 emissions + risk data.

Public Equities models:

Model A = Historical Emissions used along-side company financials, to produce “nowcast” emissions.

Model B = Company financials used to produce emissions

Model Performance Summary

Metrics Explained:

• MDAPE (Median Absolute Percentage Error): Measures the median prediction error percentage.
• MDAPE Noisy: Same metric but with ~30% random noise added to the input dataset.
• R² (R-squared): Indicates how well the model’s predictions align with the reported emissions (higher is better).

Scope 1 A Model

• Old Model:
  • MDAPE Valid: 12.061%
  • MDAPE Noisy: 19.039%
  • R²: 0.968
• New Model:
  • MDAPE Valid: 10.97%
  • MDAPE Noisy: 37.38%
  • R²: 0.978

Scope 1 B Model

• Old Model:
  • MDAPE Valid: 48.988%
  • MDAPE Noisy: 55.051%
  • R²: 0.73
• New Model:
  • MDAPE Valid: 43.95%
  • MDAPE Noisy: 60.65%
  • R²: 0.724
• Notes:
  • Our review found that the ML model was unable to accurately predict entities in the ‘Integrated Oil’ FactSet industry with low or missing NPPE, revenue or employee data. This occurred because the training data contained no reported emission data in any Integrated oil companies with NPPE or revenue below $10M. The affected entities were not included in this update.

Scope 2 A Model

• Old Model:
  • MDAPE Valid: 16.387%
  • MDAPE Noisy: 22.163%
  • R²: 0.964
• New Model:
  • MDAPE Valid: 12.825%
  • MDAPE Noisy: 21.035%
  • R²: 0.965

Scope 2 B Model

• Old Model:
  • MDAPE Valid: 35.727%
  • MDAPE Noisy: 46.489%
  • R²: 0.853
• New Model:
  • MDAPE Valid: 53.841%
  • MDAPE Noisy: 59.33%
  • R²: 0.428

Scope 3 A Model

• Old Model:
  • MDAPE Valid: 20.34%
  • MDAPE Noisy: 24.169%
  • R²: 0.895
• New Model:
  • MDAPE Valid: 20.027%
  • MDAPE Noisy: 28.254%
  • R²: 0.606

Scope 3 B Model

• Old Model:
  • MDAPE Valid: 55.246%
  • MDAPE Noisy: 57.212%
  • R²: 0.556
• New Model:
  • MDAPE Valid: 54.783%
  • MDAPE Noisy: 64.475%
  • R²: 0.766

Public equity Material Industry Intensity Factors for Scope 3 have been updated based on new reported data in these industries. Small percentage shifts were observed, but no material changes.

Private Equity models (minimal financial inputs required)

Private Equity has a new method which will preferentially leverage revenue, and fall back to Total Assets if no revenue data is available, with corresponding modelling error-rate.

Scope 1 — Revenue

• MDAPE Valid: 71.993%
• MDAPE Noisy: 71.866%
• R²: 0.239

Scope 1 — Total Assets

• MDAPE Valid: 78.303%
• MDAPE Noisy: 78.45%
• R²: 0.244

Old Scope 1 (Revenue)

• MDAPE Valid: 70.815%
• MDAPE Noisy: 70.998%
• R²: 0.402

Scope 2 — Revenue

• MDAPE Valid: 67.447%
• MDAPE Noisy: 67.951%
• R²: 0.724

Scope 2 — Total Assets

• MDAPE Valid: 77.264%
• MDAPE Noisy: 77.673%
• R²: 0.231

Old Scope 2 (Revenue)

• MDAPE Valid: 71.103%
• MDAPE Noisy: 72.639%
• R²: 0.215

Scope 3 — Revenue

• MDAPE Valid: 72.7%
• MDAPE Noisy: 74.19%
• R²: 0.319

Scope 3 — Total Assets

• MDAPE Valid: 78.195%
• MDAPE Noisy: 80.39%
• R²: 0.309

Old Scope 3 (Revenue)

• MDAPE Valid: 68.903%
• MDAPE Noisy: 70.221%
• R²: 0.174

‍