Predicting New-Vehicle CO₂ Emissions from Manufacturer Specifications: A Feature-Tier Ladder and Fuel-Type Conditional Study on Canadian EnerGuide Ratings
Year: 2026
Type: article
Abstract: Public vehicle-rating data can appear to support nearly perfect carbon-dioxide-emission prediction, although the apparent accuracy may depend more on the information supplied than on the selected algorithm. This study analysed 10,059 Canadian EnerGuide light-duty vehicle configurations from model years 2015–2024. Predictors were organized into three availability tiers: engine size and cylinders; complete manufacturer-sheet specifications; and certified city, highway, and combined fuel consumption. Ordinary least squares, random forest, and gradient boosting were compared on identical shuffled five-fold splits. The Tier 2 random forest was the strongest specification-only model, with mean R² of 0.898, mean absolute error of 13.38 g/km, and root mean squared error of 19.29 g/km. Tier 3 raised R² to 0.996–0.998, but fuel-conditional regressions produced slopes of 23.41, 23.33, 16.67, and 26.82 g/km per L/100 km for regular gasoline, premium gasoline, 85% ethanol fuel (E85), and diesel. These values closely matched official combustion-factor slopes, identifying certified fuel consumption as a target proxy rather than an ordinary pre-certification predictor. Permutation importance and vehicle-bucket analyses further showed that specification effects varied across cars, sport utility vehicles, pickups, and other vehicles. Separating transmission type from gear count allowed both components to enter the models. The results demonstrate that feature provenance matters more than small differences among common regression algorithms.
Source: Theoretical and Natural Science
Author: Jiaxin Fei
Institution: Xi’an Jiaotong-Liverpool University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
SDG: Climate action
Open Access status: hybrid
OpenAlex ID: W7223001613