AI Model Reads Lead-Lag Signals to Sharpen Carbon Price Forecasts

Carbon markets are notoriously difficult to predict. Prices in emissions trading systems do not simply react to the latest headline; they drift, overshoot, and respond to energy markets, financial conditions, and international carbon trading signals at different speeds, sometimes hours or days apart. A new study published in Complex & Intelligent Systems by Pingrang Wang, Xi Qiao, Haiyang Liu, and Xiang Wang of Zhengzhou University of Aeronautics tackles this asynchronous behavior head-on with a machine learning framework that explicitly encodes the lead-lag relationships between the many factors that drive carbon prices. The result, tested on two of China’s pilot carbon markets, beat a battery of competing methods across nearly every forecast setting the researchers examined.

The framework, called TLCC-DyG-STFusion, is built on an observation that most graph-based forecasting models have overlooked. When researchers model markets as networks, they typically connect variables that are correlated at the same moment in time. But in a carbon market, the price of electricity today may influence the carbon price tomorrow, while international allowance prices may exert their influence only after a longer delay. Existing methods, the authors argue, rarely encode these lead-lag structures as directed priors while still retaining the flexibility to adapt to each individual sample. The new framework is designed to capture both the rigid, statistically grounded timing relationships and the fluid, sample-specific associations that shift as market conditions change.

The technical core of the approach is a dual-stream architecture. In the first stream, time-lagged cross-correlation analysis, a statistical technique that measures how strongly two signals are related when one is shifted forward or backward in time, is used to construct a sparse, directed association prior. This prior acts as a kind of map of who influences whom and with what delay, distilled from the historical behavior of the market. In the second stream, the model builds a dynamic graph at the level of each individual sample, allowing the network of relationships to flex and reorganize as conditions evolve. A learnable coefficient then fuses these two views together before the combined representation passes through a spatiotemporal encoding stage that extracts patterns across both variables and time.

This fusion strategy addresses a fundamental tension in graph-based forecasting. A purely statistical prior is stable and interpretable but can become stale when the market regime shifts. A purely dynamic graph is adaptive but risks chasing noise in the data, especially in financial time series where spurious correlations abound. By blending the two with a coefficient that the model itself learns during training, the framework hedges its bets, leaning on the prior when the dynamic graph is unreliable and trusting the data-driven graph when it carries genuine, current information about the market state.

The evaluation was deliberately demanding. The researchers tested the framework on the Hubei and Guangdong pilot carbon markets, two regional emissions trading systems in China that differ in structure, liquidity, and the industrial composition of their participants. Rather than relying on a single train-test split, they used 68 rolling test blocks, a design that repeatedly re-estimates and re-tests the model as it moves through time, mimicking how a forecaster would actually deploy the system in practice. They evaluated direct forecasts at one, three, and five steps ahead, covering both short-term and medium-horizon prediction tasks.

The headline result is striking: the framework achieved the lowest five-seed mean root mean squared error in all six market-horizon settings, meaning it won in both markets at every forecast horizon tested. Averaging over five random seeds guards against the possibility that a single lucky initialization drove the results, a common pitfall in deep learning benchmarks. To assess whether the improvements were statistically meaningful rather than artifacts of chance, the authors applied Diebold-Mariano tests, the standard tool for comparing forecast accuracy, corrected using the Harvey-Leybourne-Newbold adjustment for small samples. After applying a Holm adjustment to control for multiple comparisons, the results remained significant in 56 of 60 baseline comparisons.

The authors are candid about the four comparisons where adjusted significance was not established, a level of transparency that is still uncommon in the machine learning forecasting literature. Configuration-level ablation studies, in which components of the model are systematically removed, favored the full model in all six settings, suggesting that both the prior-graph stream and the dynamic-graph stream contribute to performance. Single-seed hyperparameter sweeps showed relative changes in root mean squared error ranging from a 4.65 percent improvement to a 2.13 percent degradation, indicating that the framework’s advantage is reasonably robust to the choice of tuning settings rather than dependent on a narrowly optimized configuration.

Perhaps the most scientifically interesting part of the study is what the authors found when they examined the prior graphs themselves. Rather than treating the learned associations as a black box, they independently recomputed the prior-graph evidence and measured how stable it was across adjacent time windows. The results showed high rank continuity, with Spearman correlation means of 0.858 in the Hubei market and 0.929 in the Guangdong market, meaning the ordering of the strongest lead-lag relationships changed little from one window to the next. At the same time, the composition of factors over longer spans did change, suggesting that the market’s influence structure is stable in the short run but evolves over longer periods.

This pattern led the researchers to a carefully hedged interpretation they describe as stable-but-adaptive. Crucially, they explicitly decline to claim that the prior graphs capture causal relationships between variables. Correlation with a time lag, however suggestive, is not causation, and the authors’ willingness to delimit the interpretive scope of their method stands out in a field often criticized for overclaiming. The finding has practical implications nonetheless: if the lead-lag structure of a carbon market is sufficiently stable across adjacent windows, forecasters can rely on periodically updated priors without needing to rebuild the entire association structure from scratch every day.

For policymakers, market participants, and researchers watching the rapid expansion of carbon pricing around the world, the study offers a template for how modern graph neural networks can be adapted to the peculiar rhythms of environmental finance. China’s national emissions trading system, built on lessons from regional pilots like Hubei and Guangdong, continues to grow in importance as the country pursues its climate targets, and accurate price forecasts support everything from corporate compliance planning to the design of market-stabilizing interventions. The framework’s demonstrated cross-market competitiveness, achieving top accuracy with the same architecture in two structurally different markets, suggests the approach may generalize beyond the pilots studied here. As the authors note, the work demonstrates broad competitiveness while honestly marking the boundaries of what the evidence supports, a combination that should make it a useful reference point for the next generation of carbon market forecasting models.

Subject of Research: Machine learning-based carbon price forecasting using time-lagged cross-correlation dynamic graphs and spatiotemporal fusion

Article Title: Carbon price forecasting based on time-lagged cross-correlation dynamic graphs and a dual-stream spatiotemporal fusion network

Article References: Carbon price forecasting based on time-lagged cross-correlation dynamic graphs and a dual-stream spatiotemporal fusion network. (n.d.). https://doi.org/10.1007/s40747-026-02550-7

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02550-7

Keywords: carbon price forecasting, graph neural network, time-lagged cross-correlation, spatiotemporal fusion, dual-stream architecture, dynamic graphs, emissions trading, China carbon markets, Hubei pilot market, Guangdong pilot market, Diebold-Mariano test, forecast evaluation