Introduction
Strategic foresight is essential for policymakers navigating a world defined by Volatility, Uncertainty, Complexity, and Ambiguity (VUCA). While scenario planning is a staple for preparing for potential future scenarios, it remains a largely manual, expert-driven process. This paper introduces a novel AI-assisted framework that combines Large Language Models (LLMs) and Graph Neural Networks (GNNs) to extract and predict causal relationships from news data, offering a data-driven approach to forecasting how complex global events might influence oil prices.
The Gap in Existing Approaches
Existing AI applications in strategic foresight often focus on automating routine tasks like data scanning or topic modeling. However, there is a significant research gap in modeling causality—understanding not just that two events occur together, but that one directly causes the other. The authors aim to move beyond simple trend analysis to predict future causal relationships grounded in historical facts.
A Multi-Step Predictive Pipeline
The researchers developed a comprehensive pipeline to transform raw news into a predictive causality graph:
- Data Collection: They utilized EventRegistry, analyzing 5,486 media events (clusters of news articles) related to the "Price of Oil." The data spanned specific geopolitical windows (2015, 2020, 2022, and 2023), with contextual anchors like the Russo-Ukrainian War and U.S. Energy Information Administration reports ensuring relevance to major market shifts.
- Causal Extraction: Using ChatGPT 3.5, the team processed news summaries to identify "cause" and "effect" entities, along with their locations.
- Entity Resolution: To prevent graph clutter from different names for the same concept, text strings were mapped to standardized Wikidata concepts, creating a cleaner, more robust causality graph.
- Graph Learning: The core predictive engine relies on GraphSAGE, a GNN designed for link prediction. Each entity (node) was represented by an average word2vec embedding of its description, ensuring semantic similarity (e.g., "crude oil" and "petroleum" remain close in vector space). The GNN was then trained to predict the likelihood of a causal edge existing between two nodes based on their features and local graph structure.
Key Findings and Performance
The model was trained on historical data from 2015 and 2020 and tested on "future" events from the first quarters of 2022 and 2023.
- Quantitative Metrics: When predicting causal relationships over 2,000 pairs of entities, the model achieved a precision of 0.4905, a recall of 0.6875, and an F1 score of 0.5725. This indicates the model correctly predicted a causal link nearly half the time.
- Qualitative Success: The system successfully generated plausible causality chains, such as: OPEC → stock market index → currency → fuel pricing → Russo-Ukrainian war → energy industry.
- LLMs as Extractors: The study demonstrates that while the AI community debates whether LLMs can inherently reason causally, they are highly effective at extracting causal claims made by human journalists.
The "Expert-in-the-Loop" Vision and Future Work
The authors do not propose replacing human futurists with AI. Instead, they envision an "expert-in-the-loop" system where AI predicts potential links, and domain experts validate and define the specific nature of those relationships, making the foresight process faster and more evidence-based.
Limitations and Next Steps: Current challenges include generating long, deep, and highly accurate causality chains, as well as maintaining the "cleanliness" of the underlying causality graph. Future work will focus on improving entity matching, ranking predicted chains based on complex criteria, and ultimately using LLMs to translate these predicted causal chains into full, readable narrative scenarios for policymakers.
Summary updated