Optimization of CO₂ Emission Taxation in Switzerland Using Machine Learning

Abstract

This research aims to improve the taxation of CO₂ emissions from internal combustion engine passenger cars using a Machine Learning (ML) approach. The current tax system in Switzerland relies heavily on WLTP type-approval values, which often differ from real-world fuel consumption and emissions, limiting its effectiveness.

Using European OBFCM data and the citiwatts.eu dataset [1], the study develops predictive models to estimate actual fuel consumption and CO₂ emissions based on technical vehicle characteristics such as mass, engine power, and displacement. Ensemble algorithms, including LightGBM and Random Forest, demonstrate strong predictive accuracy, showing that ML can reliably approximate real driving emissions.

To validate the model, a qualitative assessment was conducted through a Delphi study with mobility experts, surveys of Swiss corporate fleet managers, and interviews with politicians from multiple parties. This ensured that the findings are grounded in practical, political, and economic realities.

The results indicate the need to revise local and national vehicle taxation to better reflect real emissions and to more rigorously assess plug-in hybrid vehicles, whose environmental impact is often underestimated. The research proposes a differentiated, science-based tax framework aligned with actual vehicle performance, improving both environmental effectiveness and fiscal fairness.

It also recommends enhancing consumer awareness and replacing the A–G energy labeling scale with more precise, continuous indicators.