A machine learning approach for scope-based forecasting of carbon-credit generation in voluntary carbon markets
Аннотация
Reliable carbon-credit supply projections are important for supporting climate policy, market planning, and long-term investment in climate-mitigation strategies. Forecasting carbon-credit generation, however, remains challenging because of substantial heterogeneity in project characteristics, differences in project lifecycles, and scope-specific temporal patterns. To address these challenges, this study proposes a scope-wise machine-learning framework based on the Extra Trees (ET) regression algorithm for project-level forecasting of carbon-credit generation in voluntary carbon markets. The proposed framework integrates data preprocessing, temporal panel construction, temporal and contextual feature engineering, scope-specific modelling, recursive multistep forecasting, and historical backtesting. Independent forecasting models were developed for five project scopes—Agriculture, Chemical Processes, Household & Community, Industrial & Commercial, and Renewable Energy—using a comprehensive database of historical project-level carbon-credit issuance records. The feature set incorporates lag variables, rolling averages, cumulative credits, project-level statistical descriptors, and contextual variables to capture temporal dependencies and heterogeneous project behaviour. Gradient boosting (GB) regression was used as a baseline under the same feature and evaluation settings to benchmark the performance of the proposed Extra Trees models. The Extra Trees models achieved log-scale R 2 values ranging from 72.8 to 82.1% across the five project scopes. Long-term recursive forecasting for 2025–2050 estimates approximately 2.28 billion carbon credits from 2,653 projects identified as eligible in the baseline registry, with Renewable Energy contributing the largest share of the projected credits. This projection is conditional on the projects included in the baseline registry and does not account for future project entrants or the expiry of project crediting periods; therefore, it should not be interpreted as an unconditional estimate of market-wide carbon-credit supply through 2050. The framework relies primarily on historical registry information and does not explicitly incorporate external drivers such as carbon-credit prices, policy changes, or macroeconomic conditions.