<![CDATA[Environmental commodity markets (carbon allowances, electricity, natural gas, and renewable energy instruments) are, at their root, earth system markets: their price-generating processes are as much a product of atmospheric circulation, hydrological regimes, and climate teleconnections as of supply-demand fundamentals or regulatory signals. Despite this physical reality, the machine learning forecasting literature has treated these markets primarily as benchmarking arenas, decoupling predictive architectures from the physical processes that drive price-relevant forcing. This systematic review follows PRISMA 2020 guidelines on a corpus of 413 peer-reviewed studies drawn from 652 Scopus records (2010–2026) and examines how earth system variables, including temperature anomalies, precipitation regimes, wind resource indices, atmospheric pollution metrics, and climate teleconnections such as ENSO and NAO, have been integrated into ML-based environmental commodity price models, and evaluates the evidence for whether physics-informed feature engineering confers measurable accuracy advantages over purely data-driven approaches. Bibliometric analysis reveals rapid field expansion (71.6% of publications in 2021–2026), geographic concentration in Chinese ETS research (?68% of high-impact output), and methodological dominance of hybrid decomposition-deep learning architectures. Models incorporating earth system variables consistently outperform endogenous-only ML architectures by 12–35% on MAPE, yet fewer than 30% of corpus studies include any physical predictor and climate teleconnection indices appear in under 4% of studies despite their relevance at energy market planning horizons. Six research priorities are identified, centred on numerical weather prediction ensemble integration, cross-climate-regime validation, and probabilistic forecasting grounded in physical uncertainty quantification.]]>
Copyrights © 2026
