Science and Methodology

Science and Methodology

Scientific rigour in the service of geopolitical anticipation.

Why a scientific model?

Georisk-AI does more than aggregate indicators. Our anticipation engine is built on a calibrated statistical model, validated on two decades of historical data (2000-2021).

While traditional databases merely photograph the present, our Theory-Driven Prediction model estimates the real probability of a strategic inflection point over a 36-month horizon.

This page sets out the principles of our scientific approach. To protect the intellectual property of the model, certain technical parameters remain confidential.

Model Performance

AUC (Area Under Curve)

0.85

recent decade, 2010-2021

The model's ability to tell apart the countries that will break down from those that will hold, measured through strict retrospective testing on data it had never seen. Over the full 2000-2021 period, the AUC reaches 0.81. Over the recent decade, 2010-2021, it rises to 0.85. For reference, 0.50 is equivalent to a coin toss, and 0.80 is already considered excellent in quantitative political science.

Calibration score (Brier)

0.0559

Measures the accuracy of the probabilities the model announces, not just how it ranks countries. The lower the value, the better. At 0.0559, the model's probabilities track closely what actually happens.

Comparative benchmarks

Random (chance)0.50
Typical academic statistical models0.65-0.75
Georisk-AI (walk-forward validation)0.81 · 0.85 (2010-2021)

An AUC of 0.85 means that in roughly nine cases out of ten, the model assigns a higher risk to the country that will break down than to the one that will hold.

What the model saw coming

Cases verified in retrospective validation.

  • Turkey

    Attempted coup d'état · July 2016

    Global rank #13/151 (2015) · Flagged 7-19 months ahead

  • Sudan

    Fall of the al-Bashir regime · April 2019

    Global rank #8/161 (2018) · Flagged 4-16 months ahead

  • Myanmar

    Coup d'état and civil war · Since February 2021

    Global rank #16/161 (2019) · Flagged 14-26 months ahead

  • Mali

    Coup d'état · August 2020

    Global rank #12/161 (2019) · Flagged 8-20 months ahead

  • Afghanistan

    Fall of Kabul to the Taliban · August 2021

    Global rank #5/160 (2020) · Flagged 8-20 months ahead

  • Sudan

    Civil war (SAF-RSF) · April 2023

    Global rank #2/149 (2021) · Flagged 16-28 months ahead

Risk scores computed by replaying history: the model only saw data available before each signal.

Methodology

Theory-Driven Prediction approach

Unlike purely empirical models (deep learning 'black boxes'), our approach combines political theory with structural statistics. Every variable in the model corresponds to a causal mechanism identified in the scientific literature (institutional instability, social fragmentation, economic constraints, etc.).

Statistical architecture

Structural logistic model, calibrated by penalised maximum likelihood. Explanatory variables are aggregated into weighted synthetic composites, which are then integrated into a final scoring function.

For intellectual property reasons, the coefficients, weights, and aggregation formulas are not published.

How the model is tested

The model is tested year after year, under the real conditions of a prediction. For each year between 2000 and 2021, it learns only from the past, with a four-year safety margin, then is asked to assess the risk of the following three years it has never seen. Its risk scores are then compared to what actually happened. More than twenty years put to this test, without ever being able to look ahead.

Data Sources

The model relies exclusively on public, recognised and reproducible sources. No private or unverifiable data enters the risk score calculation.

  • V-Dem (Varieties of Democracy)

    Democracy indices, political liberalism, institutional quality.

  • World Bank, World Development Indicators

    Structural economic indicators: GDP, trade balance, energy dependence.

  • WGI (Worldwide Governance Indicators)

    Governance, rule of law, control of corruption.

  • UCDP (Uppsala Conflict Data Program)

    Armed conflict data and political violence.

  • UNCTAD

    Trade flows and export concentration.

  • SIPRI

    Military expenditure and arms flows.

  • Polity Project

    Political regime and democratic transitions.

  • Maddison Project / Alesina et al.

    Long-run historical data (growth, ethnic fractionalisation).

  • UNDP, Human Development Report

    Human Development Index (HDI), 2023-24 Report, HDI vintage 2022. Displayed as is on the public map.

  • GDELT

    Real-time conflict signal (conflict status).

The public maps display internationally recognised reference indicators (HDI, GDP per capita, WGI), transparent and verifiable. The composite risk score, produced by the proprietary scientific model, is distinct.

Robust architecture

To ensure system rigor, our pipeline integrates three levels of data processing:

Level 1 : Upstream data validation

Each data source is validated syntactically and statistically before ingestion. Outliers and missing values are flagged and handled via documented fallback strategies.

Level 2 : Risk scoring model with fallback

The main engine is backed by a heuristic fallback system kept intact. Should the primary model become unavailable, the system can immediately switch to a prior version without service interruption.

Level 3 : Human override (IEGA researchers)

IEGA partner researchers can manually correct or validate variables, syntheses, and scenarios. A full audit trail records all modifications with timestamp and analyst identifier.

Scientific and methodological rigor

Human validation

AI-generated syntheses and scenarios are reviewed, validated or corrected by partner researchers. A 'Researcher-validated' badge marks content that has undergone human review.

IEGA commitment

The Institute of Applied Geopolitical Studies (IEGA) oversees the scientific rigour of the project. Partner researchers intervene for content validation and drafting the human notes commissioned by Premium subscribers.

Planned scientific publications

Our model results will be the subject of academic publications submitted to peer-reviewed journals.

Transparency vs. intellectual property

We publish our general methodology, our sources, and our performance metrics. However, the precise coefficients, weights, and formulas remain protected as trade secrets. This is consistent with industry practice at comparable providers (Verisk, Eurasia, Geoquant).

Limitations and Disclaimers

Our model is not infallible. No statistical model can predict a singular geopolitical event with certainty. Our risk scores and scenarios are decision-support tools. They are not financial, legal, or strategic advice. Users are solely responsible for their operational decisions.

See our Terms and Conditions of Sale for full liability limitations. CGV.

Let's discuss your use case

Looking to integrate our political risk analytics into your decision-making process? Our team is ready to help.