CausalAlpha: A Real-Time Geopolitical Risk Index from OSINT Channels for Causal Discovery in Financial Markets

Technical Summary of "CausalAlpha: A Real-Time Geopolitical Risk Index from OSINT Channels for Causal Discovery in Financial Markets"

1. Problem Statement

Traditional Geopolitical Risk (GPR) indices, such as the seminal work by Caldara and Iacoviello (2022), rely on newspaper archives. These sources suffer from systematic lags due to editorial curation, filtering, and publication cycles, creating a delay between geopolitical events and their measurement. While social media offers real-time data, its application as a financial signal—particularly from encrypted messaging platforms like Telegram where conflict monitoring often precedes traditional media—remains understudied. Furthermore, standard approaches to linking risk to markets, such as Granger causality, often fail to account for common causes, leading to spurious inferences. There is a need for a high-frequency, real-time GPR index derived from Open-Source Intelligence (OSINT) that employs rigorous causal discovery methods to map the directed dependencies between geopolitical narratives and financial market variables.

2. Methodology

The paper introduces CausalAlpha, an open-source framework that integrates Natural Language Processing (NLP) with constraint-based causal discovery.

• Data Construction:

  • OSINT Corpus: The system scrapes real-time messages from six Telegram channels (e.g., Bellingcat, WarTranslated, Reuters World) using the Telethon library. The corpus focuses on conflict monitoring and excludes pro-Kremlin or disinformation sources.
  • NLP Indicators: Instead of a single aggregate index, the framework generates five category-specific GPR indicators: Conflict, Political Instability, Energy Security, Trade Disruption, and Financial Stress.
  • Metric: For each category $c$ and day $t$, the raw indicator is the share of messages containing category-specific keywords. To preserve weekly variation while smoothing daily noise, a 7-day rolling share is computed.
  • Financial Variables: The study utilizes daily closing prices for 16 variables across four asset classes: commodities (e.g., Brent, Gold), FX rates, equity indices/sector ETFs (e.g., XLE, ITA), and credit instruments. All price series are first-differenced to ensure stationarity.

• Causal Discovery (PC Algorithm):

  • The core methodology employs the Peter–Clark (PC) algorithm (Spirtes et al., 2000) to recover the Markov equivalence class of the true causal Directed Acyclic Graph (DAG).
  • Procedure: The algorithm starts with a complete undirected graph and removes edges based on conditional independence tests (Fisher's Z-test) conditioned on subsets of variables. It then orients edges using Meek's propagation rules to identify V-structures.
  • Robustness Protocol: To ensure reliability, the authors estimate the PC algorithm across a $4 \times 3$ sweep:
    • 4 DAG Specifications: Varying the financial variable subset (Core Macro, Commodities, Credit & FX, Sector ETFs) while holding the five NLP indicators constant.
    • 3 Significance Levels: $\alpha \in \{0.05, 0.10, 0.15\}$.
  • Resampling: A block-bootstrap procedure (500 resamples, block size = 7 days) is used to quantify the sampling uncertainty of identified edges, preserving the autocorrelation structure of the rolling window data.
  • Validation: A Structural VAR (SVAR) is estimated on the core macro panel to generate impulse response functions (IRFs) and forecast error variance decompositions (FEVD), using the topological ordering derived from the PC algorithm.

3. Key Contributions

1. High-Frequency GPR Index: Construction of a novel, five-dimensional GPR index derived directly from real-time Telegram OSINT channels, offering a lower-latency alternative to newspaper-based measures.
2. Causal Discovery Application: Application of the PC algorithm to identify the directed causal structure between geopolitical uncertainty and financial markets, avoiding the spurious inference risks associated with pairwise Granger causality by conditioning on the full variable set.
3. Production System: Deployment of an open-source, automated production application (CausalAlpha) on Google Cloud Run that handles data collection, index construction, causal graph estimation, and reporting.

4. Key Results

The analysis yields specific findings regarding the causal architecture of geopolitical narratives and their transmission to markets:

• Causal Architecture of Narratives: Two edges are globally robust across all four DAG specifications at $\alpha = 0.10$:
  • Political Instability $\rightarrow$ Conflict Coverage
  • Energy Media Coverage $\rightarrow$ Conflict Coverage
  • This establishes "Conflict" as the primary causal sink of geopolitical narrative escalation in OSINT channels, suggesting that conflict coverage aggregates prior signals from political and energy dimensions.
• Market Transmission:
  • At the strictest significance level ($\alpha = 0.05$), Conflict Coverage $\rightarrow$ Energy Sector Equity Returns ($\Delta$XLE) is identified. This suggests that geopolitical escalation transmits to energy equity markets before fully propagating through the broader media cycle.
  • Emerging Market Bonds $\rightarrow$ VIX: The edge $\Delta$EM Bonds $\rightarrow$ VIX survives at $\alpha = 0.05$, indicating that sovereign bond stress acts as an early warning signal for broader market volatility.
  • Weak Daily Transmission: The SVAR analysis reveals that dynamic transmission from NLP signals to financial market prices is statistically weak at daily frequency. Impulse response functions show confidence bands including zero for most pairs, and FEVD indicates geopolitical NLP indicators explain less than 5% of the forecast variance of market prices.
• Information Ordering: The results suggest that geopolitical news signals operate primarily within the media narrative system (shaping how events are covered) rather than generating immediate, statistically detectable short-run price movements in financial markets at daily frequency.

5. Significance and Claims

The paper claims to provide a step toward real-time macrofinancial risk monitoring using OSINT. Its primary significance lies in:

• Methodological Rigor: Demonstrating that constraint-based causal discovery (PC algorithm) can recover directed dependencies in high-dimensional financial data without imposing prior ordering restrictions, unlike Structural VARs.
• Empirical Insight: Challenging the assumption that real-time OSINT news immediately drives daily asset prices. Instead, the findings suggest a hierarchy where political and energy narratives drive conflict coverage, which then influences specific sector equities (Energy) at strict significance levels, while the broader market transmission remains weak at daily frequencies.
• Practical Utility: Offering a framework for sovereign risk monitoring where systematic observation of political and energy narratives in OSINT channels may provide leading signals of conflict escalation before they appear in traditional indices.

The authors remain modest regarding the generalizability of their findings, noting that the corpus is heavily oriented toward the Russia–Ukraine conflict and that the "Financial" indicator is sparse. They explicitly state that the framework is a production step toward monitoring, with future work needed to extend analysis to lower frequencies and incorporate prediction market data as a third information layer.