Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
51,271 characters · 23 sections · 41 citation commands
CausalAlpha: A Real-Time Geopolitical Risk Index from OSINT Channels for Causal Discovery in Financial Markets
The measurement of geopolitical risk (GPR) and its transmission to financial markets has attracted growing attention following the seminal contribution of CaldaraIacoviello2022. GPR captures the risk associated with wars, terrorist acts, and tensions between states that affect the normal and peaceful course of international relations, and has traditionally been measured using newspaper-based sources. CaldaraIacoviello2022 document that stock returns experience a short-lived but significant drop in response to geopolitical escalation---an effect that varies substantially across industries, with the defense sector experiencing positive excess returns while sectors exposed to the broader economy, such as steelworks and mining, bear negative ones.
However, traditional media-based indices face a fundamental limitation: newspapers are slow, curated, and filtered through editorial processes that introduce a systematic lag between geopolitical events and their measurement. The proliferation of real-time social media channels---and in particular encrypted messaging platforms such as Telegram---has created a new layer of open-source intelligence (OSINT) that reflects geopolitical developments with minimal delay SoulaEtAl2024. A striking illustration is provided by the Russia--Ukraine conflict, where intelligence on Russian troops' border movements and military plans circulated on social media in the hours before President Putin announced the full-scale invasion Karalis2024.
This paper makes three contributions. First, we construct a novel high-frequency GPR index from Telegram OSINT channels using a keyword-based NLP pipeline updated daily, covering five geopolitical risk dimensions: conflict, political instability, energy security, financial stress, and trade disruption. Second, we apply causal discovery---specifically the PC algorithm of SpirtesEtAl2000---to identify the directed causal structure between our GPR indicators and a broad set of financial market variables spanning commodities, FX rates, equity sector ETFs, and credit instruments. Unlike Granger causality, the PC algorithm conditions independence tests on the full variable set, avoiding spurious inference due to omitted common causes. Third, we release an open-source production system (CausalAlpha) deployed on Google Cloud Run that automates data collection, index construction, causal graph estimation across multiple DAG specifications and significance levels, and weekly reporting.
Our results reveal a consistent causal architecture within the geopolitical media narrative system: political instability and energy media coverage causally precede conflict coverage across all four DAG specifications, while conflict coverage causally precedes energy sector equity returns at the strictest significance threshold ($\alpha = 0.05$). Dynamic transmission from geopolitical NLP signals to financial market prices is statistically weak at daily frequency, consistent with geopolitical information being incorporated into markets faster than it reaches the media cycle. While VerduzcoZanetti2026 exploit high-frequency oil futures responses to spikes in the newspaper-based GeoThreats index of CaldaraIacoviello2022 to identify geopolitical oil price shocks, we construct our indicators directly from real-time Telegram OSINT channels and apply causal discovery rather than a pre-specified structural VAR, recovering the full directed graph of dependencies without imposing prior ordering restrictions.
The paper proceeds as follows. Section (ref) reviews the related literature. Section (ref) describes the data and GPR index construction. Section (ref) presents the methodology. Section (ref) reports empirical results. Section (ref) discusses robustness. Section (ref) concludes.
The seminal contribution of CaldaraIacoviello2022 formalised the measurement of geopolitical risk through automated keyword-based searches of newspaper archives. Their monthly GPR index has been widely applied to study the effects of geopolitical uncertainty on macroeconomic outcomes, investment, and financial markets. The Economic Policy Uncertainty (EPU) index of BakerBloomDavis2016 follows a similar approach, measuring uncertainty about which economic policies will be implemented and when. Several subsequent contributions have extended these methodologies using topic modelling azqueta2023sources, azqueta2017developing and large language models ghomiraui, demonstrating that richer text representations can sharpen the measurement of uncertainty.
BollenEtAl2011 demonstrated that Twitter mood predicted DJIA movements up to four days in advance, establishing social media as a source of financially relevant information. More recently, large language models have been applied to extract sentiment and forward-looking signals from social media at scale. However, the use of Telegram as a geopolitical OSINT source---where real-time reporting by conflict monitors, military analysts, and investigative journalists frequently precedes traditional media coverage---remains understudied in the academic literature.
The causal discovery literature, following Pearl2009 and SpirtesEtAl2000, provides a principled framework for recovering causal structure from observational data without requiring experimental variation. Applications in economics include MonetaEtAl2013 on monetary policy transmission and HyvarinenEtAl2010 on structural VAR identification using non-Gaussianity. A key advantage of constraint-based methods such as the PC algorithm over pairwise Granger causality is that independence tests are conditioned on the full variable set, ruling out spurious causal inference due to common causes. This property is particularly valuable in our setting, where financial variables share common global risk drivers.
Our Telegram corpus is constructed from six OSINT and independent news channels selected for their coverage of geopolitical events, conflict monitoring, and international security. Channels were chosen to balance neutral open-source intelligence sources with editorially independent journalism, deliberately excluding sources identified as pro-Kremlin or disinformation vectors.\footnote{IntelSlava Z, a channel widely cited in earlier Telegram-based studies, was excluded on grounds of documented pro-Russian bias and disinformation SoulaEtAl2024.} The data collection pipeline uses the Telethon Python library to scrape channel messages in real time, stored in Google Firestore. Table (ref) describes the corpus.
Note that the corpus is heavily oriented toward the Russia--Ukraine conflict and associated geopolitical developments. Hence results should be interpreted as pertaining to a conflict-intensive OSINT sample rather than geopolitical risk in general. Channels with broader economic and financial coverage would be needed to generalise findings to non-conflict geopolitical risk dimensions such as trade policy uncertainty or sovereign debt stress.
Rather than constructing a single aggregate GPR index, CausalAlpha produces five category-specific indicators that capture distinct dimensions of geopolitical risk. For each category $c$ and day $t$, the raw indicator is defined as:
To smooth daily noise while preserving weekly variation, we compute a 7-day rolling share $\overline{\mathit{GPR}}_{c,t} = \frac{1}{7}\sum_{k=0}^{6}\widehat{\mathit{GPR}}_{c,t-k}$, which constitutes the indicator entering the causal analysis.
The five categories and their average message shares over the sample period are reported in Table (ref). The keyword dictionary is provided in Appendix (ref).
Figure (ref) plots the five CausalAlpha GPR indicators over the sample period. Several features of the series are worth noting. The Conflict indicator is the dominant signal throughout, reflecting the channel composition: Bellingcat, WarTranslated, and the Kyiv Independent focus heavily on the Russia--Ukraine conflict, which generates sustained conflict-related coverage at 55--70% of daily messages. The Trade indicator exhibits two sharp spikes --- in April 2025 following the US Liberation Day tariff announcements, and in October 2025 coinciding with China's sequential export bans on drone components and semiconductor chips --- consistent with these events generating discrete, short-lived surges in trade disruption coverage.
The most economically meaningful feature of the series is the structural break in the Energy indicator. Prior to July 2025 the indicator fluctuates at 2--5% of daily messages. From July 2025 onwards it permanently shifts to 10--20%, coinciding with the intensification of Ukrainian drone strikes that forced approximately 40% of Russia's oil refining capacity offline by October 2025 RussiaMatters2025. The November 2025 peak in the Energy indicator reflects Russia's 8 November attack on substations supplying the Khmelnytskyi and Rivne nuclear power plants ACLED2025. This step-change is not a data artefact --- it reflects a genuine structural shift in the geopolitical news cycle in which energy infrastructure became the primary contested domain of the conflict. The Financial indicator remains sparse throughout, consistent with the channel composition being oriented toward conflict monitoring rather than financial reporting.
We collect daily closing prices for 16 financial market variables from Yahoo Finance, spanning four asset classes: commodities (Brent crude oil, gold, wheat, copper, natural gas, silver), FX rates (USD/RUB, USD/TRY, USD/SAR, DXY), equity indices and sector ETFs (S&P 500, energy sector XLE, defense sector ITA, financials sector XLF), and credit instruments (EM sovereign bonds EMB, high-yield bonds HYG). The CBOE Volatility Index (VIX) serves as a direct measure of financial market uncertainty. All price-level series are first-differenced before estimation to ensure stationarity; VIX enters in levels.
The PC algorithm SpirtesEtAl2000 is a constraint-based causal discovery method that recovers the Markov equivalence class of the true causal DAG from observational data. We apply it to a variable set $\mathbf{V}$ comprising the five NLP indicators and a domain-specific subset of financial market variables (detailed below). The algorithm proceeds in two stages.
\paragraph{Stage 1---Skeleton estimation.} Starting from a complete undirected graph over $\mathbf{V}$, edges are removed between variable pairs $(X_i, X_j)$ if there exists a conditioning set $\mathbf{S} \subseteq \mathbf{V} \setminus \{X_i, X_j\}$ such that $X_i \perp\!\!\!\perp X_j \mid \mathbf{S}$. Conditional independence is tested using Fisher's Z-test, appropriate for approximately Gaussian data after first-differencing.
We note that the Financial indicator (0.2% average message share) and, to a lesser extent, the Trade indicator (4.5%) are sparse and right-skewed, which may violate the approximate Gaussianity assumption underlying Fisher's Z-test. Any edge involving the Financial variable should therefore be interpreted with additional caution. A rank-based alternative such as the kernel conditional independence test ZhangEtAl2012 would be more robust to non-Gaussianity but is computationally prohibitive at our sample size; we flag this as a direction for future work.
\paragraph{Stage 2---Orientation.} V-structures (colliders of the form $X_i \to X_k \leftarrow X_j$ with $X_i \not\!\perp\!\!\!\perp X_j \mid X_k$) are identified and remaining edges are oriented using Meek's propagation rules Meek1995, yielding a Completed Partially Directed Acyclic Graph (CPDAG).
To assess the robustness of causal structure across market domains and significance thresholds, we estimate the PC algorithm across a $4 \times 3$ sweep: four DAG specifications times three significance levels $\alpha \in \{0.05, 0.10, 0.15\}$ to report the highest level of significance, yielding 12 PC estimations per analysis run. Each DAG specification conditions on the same five NLP indicators but a different domain-specific market subset:
To quantify the sampling uncertainty of each identified edge, we implement a block bootstrap with 500 resamples per (DAG, $\alpha$) cell. Each resample is constructed from $\lceil N / 7 \rceil$ randomly drawn contiguous blocks of seven observations, preserving the autocorrelation structure induced by the 7-day rolling window of the NLP indicators. The bootstrap probability of an edge is defined as the fraction of resamples in which that directed edge appears in the PC output. Failed resamples contribute a conservative “edge absent” vote, ensuring bootstrap probabilities are not inflated by numerical failures.
To obtain dynamic impulse responses, we estimate a Structural VAR (SVAR) of order $p$ on the DAG 1 (core macro) sub-panel:
where $\mathbf{y}_t$ stacks the DAG 1 variables. The structural form is identified via a Cholesky decomposition using the recursive ordering derived from the PC-algorithm CPDAG: when a topological sort of the directed edges is available, it is used directly; otherwise we apply the fallback ordering VIX $\to$ Conflict $\to$ Brent $\to$ Gold specified in the architecture. We report cumulative impulse response functions (IRFs) with 95% Monte Carlo confidence intervals.
Table (ref) reports the directed edges satisfying our robustness criterion: present in all eligible DAG specifications at $\alpha = 0.10$. We use $\alpha = 0.05$ as a supplementary sensitivity check; results at $\alpha = 0.15$ are reported in Appendix (ref). Panel A reports three globally robust edges and Panel B reports three domain-specific edges present in exactly one eligible DAG specification.
Table (ref) reports bootstrap probabilities and sensitivity to significance level for all edges identified at $\alpha = 0.10$. Our robustness criterion requires presence in all eligible DAG specifications at $\alpha = 0.10$ and $\alpha = 0.15$; we use $\alpha = 0.05$ as a supplementary sensitivity check rather than a component of the criterion. Bootstrap probabilities range from 0.252 to 0.528, reflecting the moderate sample size of $N = 267$ observations (approximately 38 effective bootstrap blocks of size 7). The two globally robust edges in Panel A --- Political $\to$ Conflict and Energy $\to$ Conflict --- drop out at $\alpha = 0.05$, indicating that they reflect moderate rather than strong causal evidence; the cross-DAG criterion is the primary robustness device, with bootstrap probabilities serving as a secondary diagnostic.
One domain-specific edges in Panel B do survive at $\alpha = 0.05$: $\Delta$EM Bonds $\to$ VIX and a bootstrap probability of 0.330. This indicates that stress in the emerging market bonds ultimately originates imbalances in global financial markets. The edge between Conflict and $\Delta$XLE Energy carries the highest bootstrap probability 0.528 but does not survive the 0.05 significance interval. Recall that Energy Select Sector SPDR Fund (XLE) is the largest and most heavily traded exchange-traded fund focusing on the U.S. energy sector. Hence, changes in this ETF anticipate escalation of conflicts. This can be the case given that markets tend to anticipate large scale conflicts (see chadefaux2017market).
Looking at the four DAGs, two findings stand out. First, Conflict media coverage emerges as the primary causal sink in the geopolitical risk system, receiving causal inputs from Political and Energy in all four DAG specifications. The convergence of these three causal channels onto a single target suggests that conflict coverage in OSINT channels is not a primitive signal but rather an aggregation of geopolitical and financial pressures that have already manifested in other dimensions of the news cycle. A near-robust pattern --- present in three of four DAG specifications but absent in DAG 4 --- is VIX $\to$ Conflict, suggesting that financial market stress may amplify media attention to conflict. This direction should be treated with caution: it is not globally robust, fails at $\alpha = 0.05$, and instead suggests a feedback mechanism in which market stress amplifies media attention to conflict. This is consistent with CaldaraIacoviello2022, who document that geopolitical risk is partly endogenous to financial market conditions, and with BakerBloomDavis2016, who show that uncertainty indices respond to financial market volatility as well as generating it.
The finding that Political $\to$ Conflict and Energy $\to$ Conflict are also globally robust is consistent with the political economy literature: AlesinaPerotti1996 document that political instability is systematically preceded by deteriorating economic conditions and cuasing violent phenomena, and ChowdhuryEtAl2025 show that energy supply disruptions generate conflict escalation in energy-dependent regions.
Second, the most novel result of the analysis concerns the relationship between Conflict media coverage and $\Delta$XLE Energy identified in DAG 4 (Panel B, Table (ref)), which carries the highest bootstrap probability of all reported edges at 0.528. At $\alpha = 0.05$ --- the preferred orientation on statistical and economic grounds --- Conflict causally precedes $\Delta$XLE Energy, consistent with geopolitical escalation transmitting to energy sector equity returns. The bootstrap stability of this edge (0.528) indicates that the Conflict--$\Delta$XLE Energy relationship is the most resampling-stable finding in the analysis. The preferred $\alpha = 0.05$ orientation --- Conflict $\to$ $\Delta$XLE Energy --- is consistent with geopolitical escalation transmitting to energy equity markets before it fully propagates through the broader media cycle. This result is consistent with the broader price discovery literature documenting that forward-looking financial instruments incorporate information before it diffuses to the broader public Fama1970, and with VerduzcoZanetti2026, who show using a high-frequency Proxy VAR that geopolitical oil price shocks transmit most strongly to energy-intensive commodities, with European natural gas experiencing the largest price response among all commodity markets. Our results complement their finding by suggesting that the reverse channel also operates: energy equity markets are not passive recipients of geopolitical shocks but active price discovery mechanisms that lead the media narrative. This is consistent with informed trading in energy futures prior to conflict events, and with the episode documented by PolymarketIran2026, in which six wallets earned approximately \$1.2 million by correctly betting on prediction market contracts tied to Iranian military developments before the strikes appeared in news coverage. Crucially, VerduzcoZanetti2026 identify their geopolitical shocks using the newspaper-based GeoThreats index of CaldaraIacoviello2022; our OSINT-based indicators, which capture geopolitical signals with less editorial lag, nonetheless appear downstream of energy equity markets, suggesting that the chosen real-time Telegram channels do not fully close the information gap between financial markets and public media.
Figures (ref)--(ref) report the full CPDAGs for each of the four DAG specifications at the baseline significance level $\alpha = 0.10$. In each figure, solid arrows denote directed edges and dashed lines denote undirected edges whose orientation the PC algorithm cannot determine from the data alone.
In DAG 1 (Figure (ref)), the two globally robust edges --- Political $\to$ Conflict, Energy $\to$ Conflict, and the near-robust edges VIX $\to$ Conflict --- are all present, together with two domain-specific directed edges. These are $\Delta$Gold $\to$ Financial which implies that daily gold price movements causally precede financial stress media coverage, consistent with gold acting as a leading safe-haven barometer whose price reflects deteriorating financial conditions before they are reported in the news cycle BakerBloomDavis2016. $\Delta$SP500 $\to$ VIX confirms the well-established leverage effect in which equity market declines drive implied volatility Kyle1985. Energy $\to$ Financial is also identified, suggesting that energy stress narratives independently contribute to financial stress coverage beyond the conflict channel. Three undirected edges are present: Political $---$ Energy, Political $---$ Trade, and $\Delta$Gold $---$ $\Delta$SP500. The first two reflect strong associations between political instability, energy security, and trade disruption narratives whose causal direction cannot be identified from daily observational data --- a finding consistent with these dimensions of geopolitical risk being mutually reinforcing rather than sequentially ordered. The undirected $\Delta$Gold $---$ $\Delta$SP500 edge reflects the well-known negative correlation between equity returns and gold prices in risk-off environments, which is contemporaneous at daily frequency and therefore not orientable by the PC algorithm.
DAG 2 (Figure (ref)) confirms the two globally robust NLP edges --- Political $\to$ Conflict and Energy $\to$ Conflict --- and adds two notable undirected commodity pairs. $\Delta$Brent $---$ $\Delta$Wheat reflects the well-documented food-energy price co-movement ChowdhuryEtAl2025, amplified during the sample period by the Russia--Ukraine conflict's simultaneous disruption of grain and energy exports. $\Delta$Copper $---$ $\Delta$Natural Gas captures the co-movement between industrial metals and energy prices driven by common macroeconomic demand factors. Neither pair can be oriented at daily frequency, consistent with contemporaneous price discovery across commodity markets. Notably, all four commodity market variables are completely disconnected from the NLP indicators --- no directed or undirected edge connects the right-hand commodity cluster to the left-hand geopolitical cluster. This null result reflects the sparsity of the financial NLP indicator (0.2% average message share) and the absence of commodity-specific keywords in the OSINT channel composition, rather than a true absence of geopolitical transmission to commodity markets.
DAG 3 (Figure (ref)) yields one domain-specific directed edge beyond the globally robust set: $\Delta$EM Bonds $\to$ VIX, which survives at all three significance levels including $\alpha = 0.05$ (bootstrap probability 0.330), making it the most statistically conservative finding in the entire analysis. This result implies that emerging market sovereign bond price declines causally precede increases in implied equity volatility --- consistent with EM sovereign stress acting as an early warning signal of broader financial market fear, a transmission consistent with the 2022 Russia sanctions episode in which EM bond selloffs preceded the VIX spike by several days. The undirected $\Delta$EM Bonds $---$ $\Delta$HY Bonds edge reflects the high correlation between these two credit instruments, whose co-movement the algorithm detects but cannot orient. Notably, $\Delta$USD/RUB and $\Delta$USD/TRY are completely disconnected from both the NLP indicators and VIX in DAG 3 --- daily ruble and lira movements carry no incremental causal information about geopolitical media coverage beyond what is already captured by the NLP indicators. This null result is itself informative: currency markets in geopolitically exposed countries do not appear to lead the OSINT news cycle at daily frequency.
DAG 4 (Figure (ref)) yields an orientation-sensitive result between $\Delta$XLE Energy and Conflict media coverage. At $\alpha = 0.05$ --- the most statistically conservative threshold --- the PC algorithm orients the edge as Conflict $\to$ $\Delta$XLE Energy, consistent with the well-documented transmission from geopolitical escalation to energy sector equity returns CaldaraIacoviello2022, VerduzcoZanetti2026. At the looser thresholds $\alpha = 0.10$ and $\alpha = 0.15$ the orientation reverses, a pattern we attribute to reduced discriminatory power of the independence tests at larger conditioning sets rather than a genuine causal reversal. We therefore treat Conflict $\to$ $\Delta$XLE Energy as the preferred orientation, consistent with both the economic prior and the strictest identification criterion. Beyond this edge, DAG 4 confirms the globally robust findings: Political $\to$ Conflict and Energy $\to$ Conflict appear at all three significance levels, and $\Delta$SP500 $\to$ VIX is stable at $\alpha = 0.10$ and $\alpha = 0.15$. The four sector ETFs form a tightly connected undirected cluster at all significance levels, reflecting a common equity factor structure driven by a shared market beta that the CPDAG cannot orient given their high pairwise correlations --- an expected result consistent with the Arbitrage Pricing Theory Ross1976.
To quantify the dynamic transmission between geopolitical risk indicators and financial market variables, we estimate a Structural VAR of order $p = 1$ --- selected by AIC and BIC unanimously --- on the DAG 1 core macro panel. The Cholesky ordering follows the topological sort of the directed edges identified in DAG 1: $\Delta$SP500 $\to$ $\Delta$Gold $\to$ Political $\to$ Energy $\to$ Trade $\to$ VIX $\to$ Conflict $\to$ Financial $\to$ $\Delta$Brent Oil, placing the most exogenous variables first.
Figure (ref) reports impulse response functions for the responses of market variables (VIX, $\Delta$Brent Oil, $\Delta$Gold, $\Delta$SP500) to one-standard-deviation shocks in the geopolitical NLP indicators (Political, Energy, Conflict), with 95% Monte Carlo confidence intervals based on 1,000 replications. The predominant finding is that confidence bands are wide and include zero for virtually all geopolitical shock--market response pairs, indicating that the daily transmission from NLP geopolitical signals to financial market prices is statistically weak at this frequency and horizon. This result is consistent with the DAG analysis, where no directed edge from NLP indicators to market price series survived the robustness criteria --- geopolitical signals causally precede other NLP dimensions (e.g., Conflict) but not market prices directly.
The one exception is the Conflict $\to$ VIX response, which is negative and partially excludes zero at short horizons (days 1--3). This suggests that a shock to conflict media coverage is associated with a modest decrease in implied volatility --- a counterintuitive result that may reflect mean reversion dynamics in VIX following conflict escalation episodes, or the Cholesky ordering placing Conflict upstream of VIX which restricts the contemporaneous response.
Figure (ref) reports the forecast error variance decompositions.
The FEVD confirms the limited explanatory power of geopolitical NLP indicators for market price dynamics at daily frequency. Own-variance dominates forecast error variance for all nine variables throughout the 20-day horizon, with the geopolitical NLP indicators explaining less than 5% of the forecast variance of VIX, $\Delta$Brent Oil, $\Delta$Gold, and $\Delta$SP500 at all horizons. The Trade indicator exhibits a modest contribution to VIX variance decomposition beyond day 10, consistent with the directed Trade $\to$ VIX edge identified in the DAG analysis. These results reinforce the conclusion that geopolitical news signals operate primarily within the media narrative system --- shaping how conflict, political instability, and energy stress are covered --- rather than generating statistically detectable short-run price movements in financial markets at daily frequency. This motivates the extension to weekly or monthly aggregation as a direction for future work, where lower-frequency transmission channels may be more precisely identified.
We extend the CausalAlpha framework using a deployed prediction-market data layer based on Polymarket, the world's largest decentralised prediction market. For a curated set of liquid event contracts mapped to our five geopolitical risk categories --- covering conflict escalation, energy supply disruption, political regime change, financial sanctions, and trade restriction events --- we capture daily point-in-time implied probabilities, 24-hour trading volume, and period-over-period price changes, applying the quality filters of WolfersZitzewitz2004: minimum 24-hour trading volume of \$250,000, bid-ask spread below 5 cents on a \$1.00 contract, and at least 14 days to contract resolution.
This prediction-market layer provides a third information signal distinct from both the OSINT NLP indicators and asset prices: a forward-looking, market-implied probability series reflecting the aggregated beliefs of financially incentivised traders. Placing this series in the DAG allows us to test directly whether prediction market probabilities lead the OSINT news cycle --- sharpening the paper's central finding on information ordering --- and to locate prediction markets in the causal ordering relative to both media coverage and asset prices.
The episode documented by PolymarketIran2026 is illustrative: six wallets earned approximately \$1.2 million by correctly betting on prediction market contracts tied to Iranian military developments before the strikes appeared in news coverage, suggesting that prediction markets may incorporate non-public information faster than both OSINT channels and traditional media. If this ordering is systematic, a three-layer causal structure --- NLP indicators $\to$ prediction market probabilities $\to$ asset prices --- would represent the full information chain from geopolitical events to financial markets. We test this hypothesis and report results in ongoing work.
We assess the stability of our findings across three dimensions. First, we re-estimate all four DAGs at $\alpha \in \{0.05, 0.10, 0.15\}$. The two globally robust edges --- Political $\to$ Conflict and Energy $\to$ Conflict --- are present in all four DAG specifications at $\alpha = 0.10$ and $\alpha = 0.15$, but drop out at $\alpha = 0.05$, reflecting the limited power of independence tests at the effective sample size of approximately 38 bootstrap blocks. Second, block-bootstrap resampling (500 resamples, block size 7 days) yields probabilities of 0.330 and 0.252 for these two edges respectively, confirming moderate but not strong resampling stability. The domain-specific edge $\Delta$EM Bonds $\to$ VIX is the most statistically conservative finding, surviving at $\alpha = 0.05$ with a bootstrap probability of 0.330. Third, the multi-DAG cross-specification criterion --- requiring each edge to appear in every eligible DAG --- confirms that the globally robust edges are not artefacts of a particular asset class variable set. Complete edge lists at each significance level are reported in Appendix (ref).
A comparison of the CausalAlpha Conflict indicator with the CaldaraIacoviello2022 monthly GPR index is left for future work, as the CausalAlpha sample (April 2025--April 2026) only partially overlaps with the period for which the GPR index provides reliable monthly variation. We note that the two indicators are conceptually distinct --- the GPR index aggregates newspaper coverage of geopolitical risk broadly defined, while CausalAlpha focuses specifically on OSINT Telegram channels with a conflict-monitoring orientation --- and a meaningful comparison would require a longer overlapping sample.
This paper introduces CausalAlpha, an open-source framework for real-time geopolitical risk measurement and causal analysis of financial market dynamics. Our Telegram-based GPR index provides a high-frequency alternative to newspaper-based measures, decomposing geopolitical signals into five thematically distinct indicators across six OSINT channels. The application of the PC algorithm across four DAG specifications and three significance levels yields a consistent and economically interpretable causal architecture.
Two findings emerge as globally robust. Political instability and energy media coverage independently and causally precede conflict coverage in all four DAG specifications, establishing conflict as the primary causal sink of geopolitical narrative escalation in real-time conflict-oriented OSINT channels. At the strictest significance level ($\alpha = 0.05$), conflict coverage causally precedes $\Delta$XLE energy sector returns, consistent with geopolitical escalation transmitting to energy equity markets CaldaraIacoviello2022. The SVAR analysis indicates that transmission from geopolitical NLP signals to financial market prices is statistically weak at daily frequency, suggesting that geopolitical news signals operate primarily within the media narrative system rather than generating detectable short-run price movements.
These findings have practical implications for sovereign risk monitoring and early-warning systems: systematic monitoring of political and energy narratives in OSINT channels may provide leading signals of conflict escalation before it appears in traditional geopolitical risk indices. Future work will extend the analysis to lower frequencies where transmission to financial markets may be more precisely identified, incorporate prediction market probabilities as a third information layer between media narratives and asset prices, and apply rolling-window estimation to assess causal structure stability across geopolitical regimes.
\addcontentsline{toc}{subsection}{Appendix A: GPR Keyword Dictionary}
\addcontentsline{toc}{subsection}{Appendix B: Alpha Sweep Supplementary Results}
Tables (ref)--(ref) report the complete directed edge lists for each DAG specification at $\alpha \in \{0.05, 0.10, 0.15\}$.