Abstract
This article describes the development and deployment of the World Stability Intelligence (WSI) system — a machine learning-driven geopolitical risk monitoring platform that continuously tracks 87 countries across three risk dimensions: war risk (45%), political risk (35%), and economic risk (20%). Drawing on an ML-enhanced heuristic prediction framework (HPF-P), the system generates normalized stability scores mapped to five categories from Stable to Critical. This work discusses the methodology, current findings, and the epistemic limits of AI-based political forecasting, while making the research tools and API publicly available for researchers, journalists, and policymakers.
Keywords: geopolitical risk, machine learning, conflict prediction, state fragility, early warning systems, World Stability Intelligence, HPF-P
1. The Moment Before
On April 15, 2023, fighting broke out in Khartoum. Within forty-eight hours, Sudan had split into two armed factions — the Sudanese Armed Forces and the Rapid Support Forces — and the capital was in flames. The world called it sudden. The data did not.
Months before the first shot, the indicators were accumulating. The RSF commander had been consolidating paramilitary strength outside normal command structures. The transitional government’s legitimacy had been deteriorating since the 2021 coup. Inflation had eroded purchasing power to a breaking point. International debt negotiations had stalled. The civilian-military integration talks, hailed as progress in 2022, were quietly collapsing.
None of this was secret. It was simply scattered across too many sources, too many agencies, too many silos for any single analyst to synthesize in time. Intelligence agencies do not share their assessments. Financial markets were priced for cautious optimism until the morning artillery started. The academic literature had identified Sudan as a fragile state for years, but fragility studies are not updated in real time. They are published.
I remember watching the news that April morning with a particular kind of frustration. Not surprise at the violence — but at the gap. The gap between what the data showed and what decision-makers had in hand when they needed it most.
That gap is what we built WSI to close.
2. The Problem With Predicting What Humans Do to Each Other
Geopolitical risk assessment is one of the oldest problems in statecraft and one of the most persistently unsolved ones. This is not for lack of trying.
The field has three broad traditions that have each run hard into structural limits.
The first is human intelligence. Analysts, embassies, field reporters, and think tanks produce qualitative assessments of countries and regions. These are often excellent. They are also slow, expensive, subject to cognitive biases, and systematically constrained by what information actually reaches an analyst’s desk. The United States spends over $90 billion per year on its intelligence community and still consistently misses the specific timing of major geopolitical ruptures — from the Soviet Union’s collapse to the Arab Spring to the 2021 fall of Kabul in eleven days.
The second tradition is financial pricing. Markets are supposed to aggregate all available information. In practice, political risk is priced late and in herds. Academic research has consistently documented that sovereign bond spreads and equity volatility indicators tend to reflect instability after it has materialized in headlines rather than before (Ortiz et al., 2026, arXiv:2510.12416). Markets are reactive, not predictive — useful for measuring fear, less useful for anticipating it.
The third tradition is academic modeling. The Fragile States Index, the INFORM Risk Framework, political event databases like ACLED and GDELT — these have produced genuinely important insights into what makes states vulnerable. But they share a common limitation: they are not real-time. Annual indices capture where a country stood when the data was collected. The Fragile States Index 2026 reflects conditions observed through late 2025. By the time it is published, read, cited, and acted upon, the situation has moved.
What none of these traditions provides is a system that synthesizes structural indicators in real time, across all regions simultaneously, in a form accessible to the researchers and analysts who need it. The gap between data and decision is not primarily an intelligence problem. It is an infrastructure problem.
3. What We Built
The World Stability Intelligence system emerged from a question that is deceptively simple: if you had access to all the publicly available signals about a country’s stability trajectory, what would the composite picture look like right now?
WSI currently monitors 87 countries across three risk dimensions weighted to reflect their relative predictive contribution to stability outcomes:
War Risk (45%) is the dominant driver. This dimension integrates active conflict data, territorial disputes, military mobilization signals, cross-border tension indicators, and historical conflict recurrence rates. Countries fighting wars score high here. Countries with unresolved territorial disputes and active militarization score moderately high even without active conflict.
Political Risk (35%) is the structural layer. Government legitimacy, institutional capacity, leadership transitions, democratic backsliding indicators, protest activity, and press freedom form the core of this dimension. Political risk often leads war risk: the institutional deterioration that precedes a coup or civil conflict is measurable, and it moves on a different timeline than the violence itself.
Economic Risk (20%) is the early warning signal. Inflation trajectories, debt sustainability, currency stress, unemployment, and trade dependency metrics feed this dimension. Economic risk is the canary in the coal mine: it typically deteriorates 6 to 18 months before political risk crosses critical thresholds, and political risk similarly deteriorates before the war risk dimension spikes. Monitoring economic signals is not about predicting markets. It is about detecting the kind of material desperation that produces political radicalization.
The composite score is normalized to a 0-1 range and classified into five risk categories: Critical (0.75-1.0), High (0.55-0.75), Medium (0.35-0.55), Low (0.15-0.35), and Stable (0.0-0.15).
The underlying methodology draws on HPF-P — a Heuristic Prediction Framework that combines gradient-boosted ensemble models with domain-expert heuristics for cases where quantitative data is sparse or delayed. This hybrid approach handles one of the persistent weaknesses of purely statistical conflict prediction: the rare-event problem. Major state collapses are statistically infrequent. Pure ML models tend to under-predict them unless the architecture is specifically designed to weight asymmetric outcomes. HPF-P addresses this by encoding structural red flags as rule-based priors that interact with the statistical signal.
The system runs as a live dashboard at hub.stabilarity.com/geopolitical-risk-intelligence and exposes a public API through the Stabilarity Research Hub gateway. The API is free for researchers. The GitHub repository is available at github.com/stabilarity/geo-risk-api.
4. What the Data Shows Right Now
The current WSI snapshot, as of March 2026, presents a picture of a world that is less stable than its news cycle suggests in some places, and far more stable than its historical reputation in others.
Afghanistan: 0.905. Two decades of Western investment in institutional capacity, and the country sits at the near-maximum of the risk scale. The Taliban’s return to power did not merely reverse progress — it removed the institutional scaffolding that would have needed to exist for any recovery. Economic collapse, humanitarian crisis, and complete suppression of civil society produce a score that is nearly as high as it can go.
Syria: 0.898. The civil war that began in 2011 has never truly ended. It has merely changed shape — from full-scale insurgency to a frozen conflict with live edges, ongoing population displacement, and an economy that has contracted by over 60 percent since 2010.
Palestine: 0.877. The ongoing conflict in Gaza and the deterioration of conditions in the West Bank push this score to Critical. This is not a static situation — it is one of the most rapidly evolving risk environments in the current dataset.
Yemen: 0.860. Almost a decade of civil war, famine, cholera epidemics, and infrastructure destruction have produced one of the worst humanitarian disasters of the 21st century. Yemen’s WSI score has been in Critical territory for years. The tragedy is that this trajectory was visible well in advance of the worst suffering.
Somalia: 0.871. The original case study in state fragility. Somalia has appeared at or near the top of every fragility index for three decades. The persistence of that score is itself a finding: some countries do not trend out of Critical without structural interventions that go far beyond what the international community has been willing to provide.
The other end of the spectrum is equally instructive. Norway: 0.033. New Zealand: 0.035. Switzerland: 0.035. Canada: 0.051. These countries achieve near-minimum WSI scores through a combination of factors that are individually obvious but collectively rare: strong and legitimate institutions, high income levels with low inequality, no active territorial disputes, robust press freedom, and political systems with peaceful transfer mechanisms.
Europe’s internal variance is striking. Ukraine scores 0.769 — firmly Critical, the highest in continental Europe — while Germany scores 0.063. The gap between them represents the entire range of what stable institutions can mean for a population under stress versus one operating normally.
What the economic dimension reveals that war risk alone misses is perhaps the most actionable insight in the dataset. Several countries with no active armed conflict show elevated economic risk scores tracking the trajectory of their political risk dimension with a multi-month lag. This is the early warning signal. It does not tell you that a coup will happen in October. It tells you that the material conditions for political radicalization are deteriorating right now, and that the window for preventive action is measured in months, not years.
5. Can AI Actually Predict Collapse?
The honest answer is: not precisely. But that framing misses what models like WSI are actually good for.
The literature on automated conflict prediction has made substantial progress. Neural network architectures designed specifically for spatiotemporal conflict forecasting have demonstrated meaningful predictive power at sub-national resolution (Hegre et al., 2025, arXiv:2506.14817). Machine learning approaches applied to geopolitical and geoeconomic data have shown systematic improvements over market-only benchmarks in out-of-sample forecasting (Ortiz et al., arXiv:2510.12416, 2026). The INFORM Risk framework, now the most widely adopted quantitative early warning methodology in the humanitarian sector, has demonstrated that composite structural indicators can provide meaningful lead time before crises emerge.
But prediction and detection are different things. What AI systems do well is identify when a country’s structural profile matches the historical antecedents of instability. They are better at ruling out stability — flagging that a country’s current trajectory is inconsistent with peaceful outcomes — than at specifying the form or timing of the instability that follows.
Lead time is real but variable. Economic signals typically precede political signals by 6 to 18 months. Political signals — governance deterioration, protest escalation, legitimacy crises — typically precede acute conflict by weeks to several months. This is enough time for early responders, humanitarian pre-positioning, and diplomatic engagement. It is not enough time to reverse the underlying structural conditions that produced the risk.
WSI’s scores reflect observable structural indicators. They do not capture the specific decisions of specific actors — the general who decides to move his troops, the politician who makes the incendiary speech, the economic shock that arrives from outside the system entirely. Individual agency and exogenous shocks remain fundamentally outside the reach of any model. This is why we describe WSI as a diagnostic tool, not an oracle. It tells you what the patient’s vitals look like. It does not tell you which organ will fail first, or when.
6. Open Access: Why We Made This Public
The standard model for geopolitical risk intelligence is proprietary and expensive. The major commercial vendors sell their assessments to corporations and governments for substantial fees. This is a rational business model. It is also a structural constraint on who gets to use the best available tools.
Journalists investigating state-sponsored violence. NGOs planning humanitarian operations. Academic researchers modeling conflict dynamics. Policy analysts at smaller governments that cannot afford commercial subscriptions. These are the users who arguably need geopolitical risk intelligence most urgently, and they are the ones who have historically had the least access to it.
WSI is built to change that. The dashboard at hub.stabilarity.com/geopolitical-risk-intelligence is publicly accessible without login or payment. The API is free for researchers. The source code and API documentation are open at github.com/stabilarity/geo-risk-api.
The tools that help the world understand risk should not themselves be a risk that only wealthy actors can hedge. If a researcher in Kampala can use the same risk assessment infrastructure as an analyst in London, that is a marginal improvement in global analytical capacity. At scale, those marginal improvements matter.
7. The Responsibility
I want to end with a note that is not about methodology.
The countries at the top of our Critical list are not abstract data points. Palestine 0.877. Afghanistan 0.905. Somalia 0.871. Each of those decimal places represents real populations — children who have never known stability, hospitals that have been bombed, families that have been displaced, governments that have collapsed and reconstituted and collapsed again. The purpose of tracking collapse is not to observe it from a safe distance. It is to shorten the distance between the warning and the response.
The people who can act on early warning — aid organizations, diplomatic missions, development banks, international monitoring bodies — have historically operated with fragmented information and inadequate lead time. If WSI can give even a subset of those actors a clearer picture of where conditions are deteriorating before the deterioration becomes acute, then the data serves its purpose.
The world is not short of warning signals. It is short of systems that aggregate those signals clearly and make them accessible to the people who can act on them.
That is what we built. And it is open.
References
- Ortiz, A., et al. (2026). Geopolitics, Geoeconomics and Risk: A Machine Learning Approach. arXiv:2510.12416v4. https://arxiv.org/abs/2510.12416
- Hegre, H., et al. (2025). Next-Generation Conflict Forecasting: Unleashing Predictive Patterns through Spatiotemporal Learning. arXiv:2506.14817. https://arxiv.org/abs/2506.14817
- Fund for Peace. (2026). Fragile States Index 2026. https://fragilestatesindex.org/
- Doocy, S., et al. (2025). AI-Based Approach in Early Warning Systems. arXiv:2506.18926. https://arxiv.org/abs/2506.18926
- Ivchenko, O. (2026). World Stability Intelligence: Geopolitical Risk Monitoring via ML-Enhanced Heuristic Prediction Framework. Stabilarity Research Hub. DOI: 10.5281/zenodo.18928330