How a disciplined framework translates geopolitical shocks into portfolio-level signals for oversight
Geopolitical risk is routinely discussed in investment meetings, research notes, and risk dashboards, but it remains difficult to translate into portfolio-level analysis that can be documented and defended. The practical challenge for investment teams is determining when geopolitical developments move beyond background noise and warrant formal review.
For portfolio managers and risk committees, the issue is not a lack of information, but the absence of a disciplined way to determine whether a geopolitical development is unusual, how it might transmit through a specific portfolio, and how that assessment can be explained clearly to internal stakeholders, clients, and boards.
This post presents a structured framework for addressing that challenge. It treats geopolitical risk as a measurable time series, translates statistically significant shocks into portfolio-relevant impacts using industry sensitivities, and complements those signals with governed narrative analysis designed to support human judgment.
This discussion focuses on methodology and governance rather than prediction, with a recent geopolitical shock used solely as an illustration.
Why Geopolitical Risk Is Hard to Use in Portfolios
Daily headlines, research notes, and risk dashboards all signal that “geopolitics matters,” yet they rarely answer five practical questions:
1) Is today’s news unusual?
2) Is this just background noise, or a shock that deserves attention?
3) What does it mean for this portfolio?
4) Which industries and holdings are structurally exposed, and by how much?
5) Can we show a clear, repeatable chain from the data to the decision, suitable for clients, boards, and risk committees?
We address these questions by combining:
We illustrate the approach using a real GPR spike in June 2025 and a publicly disclosed portfolio: the iShares World ex U.S. Carbon Transition Readiness Aware Active ETF (LCTD). The ETF’s responsible investment mandate is incidental. In this illustration, it simply serves as a transparent developed market equity portfolio.
Measuring the June 23 Shock
The overlay starts from a simple principle: Treat geopolitical risk as a time series. We use the daily GPR index as a single, comparable measure of geopolitical tension across time. The first step is to determine whether a given day represents an ordinary fluctuation or an extreme shock.
Full Historical Context
Over the full history of the GPR Index (mid-1980s to 2025), most observations cluster in a relatively low range, with occasional spikes around major events such as the Gulf War, 9/11, and the invasion of Ukraine. A histogram of the full series shows a heavy right tail. Empirical quantiles mark the boundaries of “unusual” risk. Exhibit 1 illustrates:
- 95th percentile around 190
- 99th percentile around 320
- 99.5th percentile around 420
Exhibit 1: Histogram of GPR Index
Any daily reading above the 99.5th percentile is classified as an “Extreme spike” and between the 99th and 99.5th percentiles as an “Elevated spike.”
As an illustration within this framework, June 23 stands out as one of the highest readings in the sample:
- GPR level at peak: approximately 542
- Percentile: 99.8% of all daily observations
- Label: Extreme spike
To provide context, we define a fixed analysis window around the peak:
- Start: June 16, 2025
- End: June 25, 2025
Within that window, the overlay treats June 23 as the shock date and the surrounding days as the buildup and immediate aftermath.
Exhibit 2: June 2025 Geopolitical Risk Spike
Time series of the GPR index highlighting the June 23 extreme spike, with the surrounding analysis window shaded.
This event provides the stress template for the rest of the analysis. The question is, “How would a portfolio like LCTD be expected to behave, conditional on a GPR shock of this magnitude and profile?”
Translating GPR into Portfolio Terms
The framework converts headline shocks into basis point risk using a deterministic two-stage process implemented in Python. First, every security in the LCTD portfolio is mapped to the Federal Reserve’s industry taxonomy. Each industry carries a pre-estimated GPR beta that summarizes how its daily returns have historically correlated with the Caldara-Iacoviello index. Second, the June 23 spike is fed through those betas. Industry scores are scaled by position weights and then summed, producing both a portfolio level impact number and a full cross section that shows which sectors drive it.
Illustrative Portfolio
We used LCTD for this illustration because it offers:
- A diversified, developed market equity portfolio
- Sector weights broadly similar to global ex US benchmarks
- A modest tilt towards lower carbon and transition ready companies
The five largest weights are HSBC at 1.9% (Banks), AML at 1.7% (Semiconductors), AstraZeneca at 1.7% (Pharma), Iberdrola at 1.4% (Utilities) and Allianz at 1.3% (Insurance). All issuer-level references that follow use these real names and weights, drawn directly from the public holdings file.
Industry Breakdown and Vulnerability
Each security is mapped to one of 12 Fed industries (e.g., machinery, computers, depository institutions). For each industry we compute:
- Portfolio weight (%)
- Estimated GPR beta (sensitivity to the GPR factor)
- Impact score for the June 23 spike, translated into basis points of expected effect on the portfolio’s return for that event
Based on the sign of the impact score and economic reasoning, industries are classified as:
- Vulnerable (expected to be hurt by the shock), or
- Resilient (expected to benefit or provide ballast).
For the June 23 spike and the LCTD portfolio, the overlay estimates:
- Total negative impact: ≈ 33.8 bps
- Total positive impact: ≈ +15.3 bps
- Net GPR impact: ≈ 18.4 bps
In other words, conditional on a shock of this severity, the portfolio is tilted modestly toward GPR-sensitive industries, with an expected drag of roughly 18 basis points compared with a GPR-neutral configuration.
The vulnerability composition is summarized as:
- 39% of portfolio weight in vulnerable industries
- 61% in non-vulnerable or resilient industries
- five of 12 industries classified as vulnerable by the model
Exhibit 3: Industry-Level GPR Impact for the June 23, 2025, Spike
Bar chart of industry impacts (in basis points) ordered from most negative to most positive, with colors indicating vulnerable vs. resilient industries.
Key observations:
- Machinery is the largest source of downside GPR exposure, with an estimated impact of about 16.5 bps, reflecting both a meaningful portfolio weight and a negative GPR beta.
- Consumer discretionary and construction materials contribute additional downside of roughly 9.9 bps and 3.4 bps, respectively.
- On the positive side, computers (+7.0 bps), foodstuff (+4.6 bps), and depository institutions (+1.6 bps) provide partial offset.
Exhibit 4: Industry Weight vs. Impact
This scatter plot of industry weight vs impact highlights that the portfolio’s single most important trade-off is between a sizeable overweight in machinery (negative) and a large allocation to banking and technology (mildly positive in this scenario).
From Spikes to Storylines
The quantitative overlay deliberately stops at the industry level. It answers, “how much” and “where,” but not “why” or “what to do.” Those questions are managed by an AI-supported narrative layer that operates on three levels, always with a human analyst in the loop.
In this illustration, the AI-supported layer follows three governed workflows:
- Geopolitical event discovery, which identifies and clusters the real-world developments behind a statistical spike.
- Economic channel mapping, which translates those events into industry-level economic effects using a constrained taxonomy.
- Stock-level prioritization, which flags individual holdings that may warrant closer review.
The design follows CFA Institute guidance on explainable AI: Models are kept separate from judgement, reasoning paths are logged, and the technology augments but never replaces human decision makers.
Geopolitical Event Discovery: “What Just Happened?”
Once the Python engine flags June 23 as a 99.8ᵗʰ-percentile spike, the first agent fans out across curated news feeds and structured data sources. Using a fixed lexicon of geopolitical themes, it hoovers up reporting for the 10-day window around the spike, drops metaphors and noise (“trade-war-of-words,” “hockey war,” etc.), and groups what remains into a handful of coherent storylines.
For the June episode three clusters emerged naturally:
- Escalation in the Middle East energy corridor: missile exchanges, tanker-rate surges, Strait-of-Hormuz coverage.
- Red-Sea shipping threats and Houthi activity: container traffic rerouting, marine-insurer premium shocks.
- US homeland-terror and cyber alerts: FBI warnings, suspected Iran-linked cyber probes of critical infrastructure.
Each cluster is returned with a two-sentence plain English summary, a severity flag, and live links to the underlying articles. Nothing about holdings or economics is inferred at this stage; the goal is simply to agree on which real-world events drove the statistical outlier.
Economic Channel Mapping: “So What?”
The second agent receives two inputs: the threat clusters above and the portfolio’s industry impact sheet. It aims to bridge the gap between geopolitics and economics by performing these three verifiable moves behind the scenes:
- Evidence synthesis: For each cluster it scrapes dedicated financial news APIs and macro datasets such as FMP for revenue by geography, company mission statement, and sanction updates. All raw snippets are stored so an auditor can trace every claim.
- Channel tagging: Using a restricted taxonomy — energy-supply risk, maritime trade disruption, and cyber security demand – a macro-confidence shock is applied to the evidence with zero shot classifiers (LLM). The mapping is deterministic: given the same evidence, the same tags appear.
- Industry linking: Tags are cross walked to the industries that already carry GPR betas. Direction and strength come from the overlay’s numbers; the agent merely narrates them. For example: The Middle East escalation maps to petroleum & natural gas, machinery, and construction-materials (higher input costs, cap-ex delays); Red-Sea trade disruption hits computers and electronics equipment via freight delays; cyber-alerts lift demand for segments of computers and communication.
To keep the workflow auditable, the agent must cite at least one piece of verifiable data for every tag it assigns. It never rewrites scores, never creates new industries, and never overrides the quant model.
Stock-Level Exposure and Priority Review
The third AI agentic workflow works at the holding level, using the industry-level signals and the portfolio holdings file, deeply investigating specific evidence from news and fundamentals for each holding in the portfolio.
It produces a prioritized watchlist of holdings with:
- Weight, industry, and role (vulnerable/resilient)
- A one-sentence rationale grounded in the earlier channels
- A recommended priority level for risk review (high/medium/low)
Table 1: Priority Holdings (LCTD) Under the June 23 Shock
In practice, an analyst or portfolio manager reviews this list, challenges the rationales, and decides whether to run scenario analysis on the most exposed names, adjust position sizes, or document the assessment and keep the positions unchanged.
Governance, Explainability, and Auditability
An overlay that links geopolitics to holdings must meet a higher bar for governance than a stand-alone risk index. Two features are central.
Python engine (deterministic):
- Spike detection and classification
- Industry betas and impact scores
- Portfolio vulnerability summary
The AI layer (decision support):
- It cannot alter impact scores or invent holdings
- Every narrative element is grounded in retrieved documents or fundamentals
- Templates and prompts enforce structured, concise rationales rather than opaque prose
This aligns with CFA Institute guidance: AI tools must be explainable, auditable, and under human oversight, not black box trade machines.
Repeatability and Documentation
To support transparency and independent scrutiny, the materials used in the June 23 illustration are publicly available. For this illustration, the public GitHub repository includes the GPR time series and spike-classification code and the portfolio holdings extract and mapping to Federal Reserve industry classifications. Jupyter notebooks recreate Exhibits 1 to 3 along with supporting diagnostics and structured, machine-readable outputs of the portfolio impact engine. JSON files cover event metadata, industry-level impacts, and vulnerability composition.
This allows readers to trace results from input data through to portfolio-level signals, adjust parameters where appropriate, and test how the framework behaves when applied to different portfolios or stress events.
In Practice
The combined overlay does not predict conflicts, nor prescribe trades. Instead, it provides a lens for incorporating geopolitical risk into portfolio oversight.
In practical terms, it allows a portfolio team to:
- Detect when geopolitical risk genuinely moves into unusual territory, rather than reacting to every headline.
- Quantify how a specific portfolio is tilted across vulnerable and resilient industries in basis point terms.
- Explain the results in plain language, connecting the numbers to geopolitical events, economic channels, and stock level exposures.
- Document a clear, auditable assessment of how the portfolio might behave under a defined stress event.
The framework is designed to inform oversight decisions, specifically enhanced monitoring, documented risk assessment, and targeted scenario analysis. It does not prescribe trades or portfolio rebalancing.
For CIOs, risk committees, and clients, this bridges the gap between “we monitor geopolitics” and “here is how this particular geopolitical shock would transmit through your holdings.”
The June 23 spike is only one episode, but it shows that mapping headlines into holdings is feasible with a disciplined combination of data, models, and carefully governed AI.
