Deep Causal Learning to Explain and Quantify the Geo-Tension's Impact on Natural Gas Demand

Philipp K. Peter, Yulin Li, Ziyue Li, Wolfgang KetterISGT EUROPE 2024 · DOI

Introduction

Predicting natural gas demand (NGD) is a complex challenge, traditionally governed by predictable cycles of weather seasonality and economic activity. However, geopolitical "shocks"—such as the Russia-Ukraine conflict—introduce extreme volatility and non-linearities that render standard forecasting models, like ARIMA or linear regression, largely ineffective. These traditional methods often assume stationarity and linear relationships, conditions that dissolve during periods of international crisis.

Researchers at the University of Cologne have addressed this gap by developing a "Deep Causal Learning" framework. This approach moves beyond simple black-box prediction to provide a causal explanation of how the war specifically impacted the German natural gas market. By combining deep learning with counterfactual modeling, the study quantifies the "treatment effect" of the conflict across different economic sectors.

Methodology: The Deep Causal Learning Pipeline

The study employs a three-pronged methodological pipeline designed to isolate the impact of geopolitical tension from other variables like temperature or price fluctuations.

1. Factor Selection via Mutual Information (MI)

To manage the high dimensionality of potential demand drivers, the researchers use Mutual Information (MI). Unlike standard correlation coefficients, MI identifies both linear and non-linear dependencies between variables. This step ensures that the models focus on the most relevant inputs—such as Heating Degree Days (HDD), gas prices (THE Price), and supply metrics—across the Residential (LDZ), Industrial (IND), and Gas Power Plant (GTP) sectors.

2. Deep Non-linear Granger Causality

While standard Granger causality tests assess whether one time series can help predict another in a linear framework, the authors utilize Long Short-Term Memory (LSTM) networks to capture non-linear temporal dependencies. By comparing the predictive accuracy of an LSTM model with and without a "War" dummy variable, and applying a Wilcoxon signed-rank test, they statistically validated that the conflict was a primary causal driver of demand shifts, rather than a mere correlation.

3. Counterfactual Intervention Study

To quantify the magnitude of the impact, the researchers generated a "business-as-usual" counterfactual. They trained models (LSTM and Facebook’s Prophet) on pre-war data (April 2010 to February 2022) to project what gas demand would have looked like had the war not occurred. By comparing these projections to the actual observed demand during the conflict period, they isolated the specific reduction in consumption attributable to the geopolitical shock.

Sectoral Insights and Results

The research reveals that the impact of the war was not uniform across the German economy, showing significant sectoral heterogeneity:

  • Industrial Sector (IND): This sector exhibited the most immediate and stable response. Industrial demand saw a systematic decline shortly after the invasion, likely driven by rapid fuel switching, production curtailments, or efficiency measures in response to soaring prices.
  • Residential Sector (LDZ): The impact on households was characterized by a "lagged" effect. While the war's influence was statistically significant (p-value of $1.1 \times 10^{-16}$), the reduction in demand only became a dominant factor in September 2022. This suggests that behavioral changes in households and the delayed pass-through of wholesale energy prices to consumers took several months to manifest.
  • Gas Power Plants (GTP): This sector proved to be the most volatile and difficult to attribute solely to the war. Demand in this sector is influenced by a complex web of factors, including the availability of French nuclear power and domestic renewable energy (wind and solar) output.

Conclusion

The study demonstrates that LSTMs generally outperform traditional statistical methods in handling the non-stationarity introduced by geopolitical crises. The "Deep Causal Learning" framework provides a robust tool for policy-makers to disentangle the effects of external shocks from seasonal trends.

While the model successfully identifies the war as a structural break, the authors note that future work is needed to further refine the mechanisms of change—specifically distinguishing between government-mandated savings and voluntary reductions. Future iterations of this research may integrate Large Language Models (LLMs) to analyze news sentiment as a real-time proxy for geopolitical tension, potentially offering even more granular insights into market disruptions.

Summary updated

BibTeX
@inproceedings{peter2024deep,
  author = {Philipp K. Peter and Yulin Li and Ziyue Li and Wolfgang Ketter},
  title = {{Deep Causal Learning to Explain and Quantify the Geo-Tension's Impact on Natural Gas Demand}},
  booktitle = {ISGT EUROPE},
  year = {2024},
  doi = {10.1109/isgteurope62998.2024.10863241},
}

Loading related content...

Figure