Pith. sign in

REVIEW 2 major objections 1 minor 27 references

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

T0 review · 2 major / 1 minor · reviewed 2026-06-27 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read CausalAlpha recovers directed causal links from Telegram OSINT showing political instability and energy coverage precede conflict coverage, which then precedes energy equity returns.

desk verdict The paper turns Telegram OSINT into five labeled GPR series, runs PC discovery across specs and bootstraps, and reports that instability and energy coverage precede conflict while conflict precedes delta XLE at alpha 0.05, with weak daily market transmission. read the letter →

arxiv 2606.07049 v1 pith:K4VHPIHO submitted 2026-06-05 econ.EM

Andres Azqueta-Gavaldon , Borja Ureta This is my paper · ORCID
classification econ.EM
keywords geopoliticalriskcausaldiscoveryOSINTPCalgorithmfinancialmarketsTelegramNLPdirectedacyclicgraph
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper develops an open-source framework called CausalAlpha to build a high-frequency geopolitical risk index by applying natural language processing to Telegram OSINT channels. It then uses the Peter-Clark algorithm to recover directed acyclic graphs linking five category-specific risk indicators to financial variables including commodity prices, equity indices, and credit instruments. Across four DAG specifications and multiple significance levels with block-bootstrap resampling, the analysis identifies robust patterns in which political instability and energy media coverage independently cause conflict coverage. At the strictest threshold, conflict coverage causes shifts in energy sector equity returns. A structural VAR confirms weak direct transmission from the signals to market prices at daily frequency, indicating that most effects remain inside the media narrative system.

What carries the argument

The Peter-Clark (PC) algorithm, which recovers the directed acyclic graph of causal dependencies among the five GPR categories and financial time series variables.

What would settle it

An independent dataset in which the edges from political instability and energy coverage to conflict coverage fail to appear consistently across bootstrap resamples, or in which conflict coverage does not precede energy equity returns at alpha = 0.05.

Watch

Extended reading notes

Core claim

By constructing five category-specific GPR indicators via NLP on real-time OSINT and estimating DAGs with the PC algorithm under multiple specifications and significance levels using 500 block-bootstrap resamples, the paper establishes that political instability and energy media coverage causally precede conflict coverage at alpha = 0.10 across all DAGs. At alpha = 0.05, conflict coverage causally precedes energy sector equity returns. The structural VAR analysis shows statistically weak dynamic transmission from the geopolitical NLP signals to financial prices.

Load-bearing premise

The PC algorithm recovers the true causal directed acyclic graph from the observed time series under assumptions of causal sufficiency, faithfulness, and no latent confounders.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces CausalAlpha, an open-source framework that extracts five category-specific geopolitical risk (GPR) indicators from Telegram OSINT channels via NLP and applies the PC algorithm to recover directed acyclic graphs (DAGs) linking these indicators to financial variables (commodity prices, equity indices, credit instruments). It reports results across four DAG specifications, three significance levels, and 500 block-bootstrap resamples. Two findings are claimed as globally robust at alpha=0.10: political instability and energy media coverage independently causally precede conflict coverage. At alpha=0.05, conflict coverage causally precedes delta XLE. A follow-on SVAR finds weak transmission from the NLP signals to market prices at daily frequency. The framework is deployed as a production application on Google Cloud Run.

Significance. If the causal claims hold under the stated assumptions, the work supplies a novel high-frequency, real-time GPR index from OSINT sources together with empirical evidence on the ordering of geopolitical narrative escalation and its transmission to energy equity returns. The open-source code, block-bootstrap robustness protocol, and deployed production system are concrete strengths that could support reproducible macrofinancial monitoring applications.

major comments (2)
  1. [Abstract] Abstract: the central claim that political instability and energy coverage 'independently and causally precede' conflict coverage (and that conflict precedes delta XLE) is obtained from the PC algorithm. PC recovers a DAG only under the causal sufficiency assumption (no latent confounders). Daily OSINT coverage variables are jointly driven by unobserved real-world geopolitical events that are not captured in the five categories; these events constitute latent common causes that can induce spurious conditional dependencies. The four DAG specifications and block-bootstrap resampling do not relax this assumption, so the recovered edges cannot be interpreted as causal precedence without additional justification or alternative methods that tolerate latent variables.
  2. [Abstract] Abstract (methods description): the manuscript states that results are 'estimated across four DAG specifications' yet provides no explicit definition of how the four specifications differ (e.g., variable inclusion, background knowledge constraints, or conditioning sets). Without this information it is impossible to assess whether the reported global robustness at alpha=0.10 is an artifact of specification choice rather than a genuine property of the data-generating process.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'causally precede' is used without reminding the reader that this interpretation is conditional on the standard PC assumptions (causal sufficiency, faithfulness, i.i.d. sampling). A brief parenthetical qualifier would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the assumptions and clarity of our causal discovery framework. We respond point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that political instability and energy coverage 'independently and causally precede' conflict coverage (and that conflict precedes delta XLE) is obtained from the PC algorithm. PC recovers a DAG only under the causal sufficiency assumption (no latent confounders). Daily OSINT coverage variables are jointly driven by unobserved real-world geopolitical events that are not captured in the five categories; these events constitute latent common causes that can induce spurious conditional dependencies. The four DAG specifications and block-bootstrap resampling do not relax this assumption, so the recovered edges cannot be interpreted as causal precedence without additional justification or alternative methods that tolerate latent variables.

    Authors: We agree that the PC algorithm requires the causal sufficiency assumption and that unobserved real-world events could induce latent confounding not addressed by our robustness checks. The multiple specifications and block-bootstrap protocol test sensitivity to observed-data variation but do not relax the no-latent-confounder assumption. In revision we will (i) state the assumption explicitly in the methods, (ii) qualify all causal-precedence language in the abstract and results as conditional on causal sufficiency, and (iii) add a limitations paragraph noting that methods tolerant of latent variables (e.g., FCI) are left for future work. These changes will be reflected in both the abstract and the main text. revision: partial

  2. Referee: [Abstract] Abstract (methods description): the manuscript states that results are 'estimated across four DAG specifications' yet provides no explicit definition of how the four specifications differ (e.g., variable inclusion, background knowledge constraints, or conditioning sets). Without this information it is impossible to assess whether the reported global robustness at alpha=0.10 is an artifact of specification choice rather than a genuine property of the data-generating process.

    Authors: We accept that the current manuscript does not define the four DAG specifications. The specifications vary by the set of background-knowledge constraints imposed on the PC search and by the precise variable subsets included in each run. In the revised version we will insert a dedicated methods subsection that enumerates each specification, lists the exact constraints and variable inclusions, and reports the corresponding adjacency matrices so that readers can replicate and evaluate the robustness claim. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; central claims are direct empirical outputs of standard PC algorithm on observed series

full rationale

The paper applies the PC algorithm (with block-bootstrap) across four DAG specifications and three alpha levels to recover edges among five GPR categories and financial variables, then reports which edges remain robust. These outputs do not reduce by construction to any fitted parameter, self-citation chain, or renamed input; the derivation chain consists of data ingestion, NLP categorization, conditional-independence testing, and robustness checks whose results are not presupposed by the method itself. No load-bearing step matches the enumerated circularity patterns. Minor post-hoc tuning risk on alpha or categories is noted but does not constitute circularity under the stated criteria.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The paper relies on the standard causal assumptions of the PC algorithm (faithfulness, causal sufficiency, no latent variables) and on the accuracy of the NLP classifier for labeling Telegram messages. No free parameters are explicitly fitted to produce the headline causal edges; the significance levels are chosen rather than estimated. No new entities are postulated.

assumptions (2)
  • domain assumption The PC algorithm recovers the true DAG under faithfulness and causal sufficiency
    Invoked when the paper interprets recovered edges as causal precedence between GPR categories and financial variables.
  • domain assumption NLP classification of Telegram messages into the five GPR categories is sufficiently accurate for causal analysis
    Required for the input time series to be valid; not tested or reported in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CausalAlpha: A Real-Time Geopolitical Risk Index from OSINT Channels for Causal Discovery in Financial Markets." pith.science (2026). https://pith.science/paper/K4VHPIHO

@misc{pith2026260607049,
  author       = {Pith},
  title        = {Pith review of: CausalAlpha: A Real-Time Geopolitical Risk Index from OSINT Channels for Causal Discovery in Financial Markets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4VHPIHO}},
  note         = {Machine review of arXiv:2606.07049}
}
read the original abstract

We introduce CausalAlpha, an open-source framework that constructs a high-frequency Geopolitical Risk (GPR) index from Telegram OSINT channels using natural language processing, and applies causal discovery methods to identify the directed causal structure between geopolitical uncertainty and financial market variables. Unlike standard sentiment indices or Granger-causality approaches, CausalAlpha employs the Peter-Clark (PC) algorithm to recover the directed acyclic graph (DAG) of causal dependencies between five category-specific GPR indicators and a set of financial variables spanning commodity prices, equity indices, and credit instruments, estimated across four DAG specifications and three significance levels with 500 block-bootstrap resamples. Two findings emerge as globally robust across all DAG specifications at alpha = 0.10: political instability and energy media coverage independently and causally precede conflict coverage, establishing conflict as the primary causal sink of geopolitical narrative escalation in real-time OSINT channels. At the strictest significance level (alpha = 0.05), conflict coverage causally precedes energy sector equity returns (delta XLE), consistent with geopolitical escalation transmitting to energy markets. A Structural VAR on the core macro panel confirms that dynamic 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. The framework is deployed as a production application on Google Cloud Run with automated data collection and index construction, representing a step toward real-time macrofinancial risk monitoring using OSINT.

Figures

Figures reproduced from arXiv: 2606.07049 by the authors.

Figure 1
Figure 1. CausalAlpha GPR indicators, daily and 7-day rolling share, April 2025–April 2026. Shaded lines show daily raw keyword shares; bold lines show 7-day rolling means (equation 1). Dashed vertical lines indicate key geopolitical events: US–Israel strikes on Iranian nuclear facilities (June 2025), Russia’s massive attack on Kyiv (September 2025), intensification of Ukrainian drone strikes on Russian energy infrastructure … view at source ↗
Figure 2
Figure 2. Causal DAG 1 — NLP indicators + core macro variables (VIX, [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Causal DAG 2 — NLP indicators + commodity markets ( [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Causal DAG 3 — NLP indicators + credit and currency markets ( [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Causal DAG 4 — NLP indicators + equity sector ETFs ( [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Market responses to geopolitical shocks — SVAR (Cholesky, 95% MC confidence [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Forecast error variance decomposition (FEVD). Ordering: [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references

  1. [1]

    American Economic Review , year =

    Caldara, Dario and Iacoviello, Matteo , title =. American Economic Review , year =

  2. [2]

    and Bloom, Nicholas and Davis, Steven J

    Baker, Scott R. and Bloom, Nicholas and Davis, Steven J. , title =. Quarterly Journal of Economics , year =

  3. [3]

    Journal of Computational Science , year =

    Bollen, Johan and Mao, Huina and Zeng, Xiaojun , title =. Journal of Computational Science , year =

  4. [4]

    Pearl, Judea , title =

  5. [5]

    Spirtes, Peter and Glymour, Clark and Scheines, Richard , title =

  6. [6]

    and Coad, Alex , title =

    Moneta, Alessio and Entner, Doris and Hoyer, Patrik O. and Coad, Alex , title =. Oxford Bulletin of Economics and Statistics , year =

  7. [7]

    Estimation of a Structural Vector Autoregression Model Using Non-

    Hyv. Estimation of a Structural Vector Autoregression Model Using Non-. Journal of Machine Learning Research , year =

  8. [8]

    Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence , year =

    Meek, Christopher , title =. Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence , year =

Show all 27 references
  1. [9]

    Economics Letters , volume=

    Developing news-based economic policy uncertainty index with unsupervised machine learning , author=. Economics Letters , volume=. 2017 , publisher=

    2017

  2. [10]

    European Economic Review , volume=

    Sources of economic policy uncertainty in the euro area , author=. European Economic Review , volume=. 2023 , publisher=

    2023

  3. [11]

    Banco de Espana working paper, N.º 2609 , year=

    RAUI: UNCERTAINTY INDICATORS BUILT WITH ARTIFICIAL INTELLIGENCE , author=. Banco de Espana working paper, N.º 2609 , year=

  4. [12]

    2024 , url =

    Soula, Emerson and Dubow, Ben and Osadchuk, Roman , title =. 2024 , url =

    2024

  5. [13]

    Georgetown Journal of International Affairs , year =

    Karalis, Magdalene , title =. Georgetown Journal of International Affairs , year =

  6. [14]

    European Economic Review , year =

    Alesina, Alberto and Perotti, Roberto , title =. European Economic Review , year =

  7. [15]

    , title =

    Hadzi-Vaskov, Metodij and Pienknagura, Samuel and Ricci, Luca A. , title =. 2021 , doi =

    2021

  8. [16]

    2021 , doi =

    Barrett, Philip and Bondar, Mariia and Chen, Sophia and Chivakul, Mali and Igan, Deniz , title =. 2021 , doi =

    2021

  9. [17]

    2026 , month =

    Verduzco-Bustos, Guillermo and Zanetti, Francesco , title =. 2026 , month =

    2026

  10. [18]

    2025 , url =

    The Russia--. 2025 , url =

    2025

  11. [19]

    2025 , url =

    Europe and. 2025 , url =

    2025

  12. [20]

    Kabir and Abdullah, Mohammad and Hossain, Md Mofazzal , title =

    Chowdhury, Mohammad Ashraful Ferdous and Hassan, M. Kabir and Abdullah, Mohammad and Hossain, Md Mofazzal , title =. Quantitative Finance and Economics , year =

  13. [21]

    , title =

    Fama, Eugene F. , title =. Journal of Finance , year =

  14. [22]

    , title =

    Kyle, Albert S. , title =. Econometrica , year =

  15. [23]

    , title =

    Ross, Stephen A. , title =. Journal of Economic Theory , year =

  16. [24]

    CoinDesk , year =

    Genç, Timur , title =. CoinDesk , year =

  17. [25]

    Journal of Peace Research , volume=

    Market anticipations of conflict onsets , author=. Journal of Peace Research , volume=. 2017 , publisher=

    2017

  18. [26]

    Kernel-based Conditional Independence Test and Application in Causal Discovery , booktitle =

    Zhang, Kun and Peters, Jonas and Janzing, Dominik and Sch\". Kernel-based Conditional Independence Test and Application in Causal Discovery , booktitle =. 2012 , pages =

    2012

  19. [27]

    Journal of Economic Perspectives , year =

    Wolfers, Justin and Zitzewitz, Eric , title =. Journal of Economic Perspectives , year =

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.