Executive summary
- The purpose of a geopolitical risk model is consistency and traceability, not prediction of specific events.
- Event data quality — definition, coverage, verification and lag — is the binding constraint on model quality.
- Dynamic specifications such as self-exciting and regime-switching models capture clustering and abrupt change better than static rates.
- Validation must rely on distributional and process-based tests, since point-event prediction is not the objective.
- Model output belongs alongside qualitative assessment, not in place of it.
Why model at all
Underwriting conflict-exposed business has always involved judgement informed by intelligence, market signals and experience. The case for adding a quantitative layer is not that models judge better. It is that models judge consistently, leave an audit trail, allow the effect of each factor to be examined, and can be applied at portfolio scale in a way that individual assessment cannot.
- Consistency: the same exposure receives the same assessment regardless of who reviews it or when.
- Traceability: the reason for a rate change can be attributed to a specific input movement.
- Portfolio scale: accumulation across hundreds of live transits cannot be assessed by narrative review.
- Capital: boards and regulators require exposure expressed in loss and capital terms.
Data sources and event classification
Public conflict-event datasets, maritime incident reporting, vessel-tracking data and open-source reporting each contribute. They differ substantially in what counts as an event, in geographic and temporal coverage, in verification standards and in reporting lag. A model built on one source and validated against another will show discrepancies that reflect the sources rather than the phenomenon.
- Adopt a single explicit event taxonomy covering incident type, actor category, target category, location precision and severity.
- Map every source into that taxonomy and record the mapping decisions, including cases where mapping is ambiguous.
- Handle duplicates across sources deliberately; the same incident reported three times will inflate intensity estimates.
- Record reporting lag by source and adjust recent-period intensity estimates for expected under-reporting.
- Maintain traffic or exposure denominators from independent data so that intensity is a rate rather than a count.
Classification is not a clerical task. Whether a drone interception counts as an attack, whether a detention counts as a security incident, and whether an unclaimed explosion is attributed to a particular actor all change the resulting parameters. These decisions should be documented and reviewed rather than resolved case by case by whoever is processing the data.
Dynamic model structures
Conflict events are not independent and identically distributed. They cluster in time, escalate and de-escalate, and shift between qualitatively different regimes. Static average rates therefore misstate risk in both directions: too high during quiet periods, far too low immediately after an escalation.
- Self-exciting point processes represent the empirical fact that one incident raises the short-term probability of further incidents.
- Regime-switching and hidden Markov structures represent discrete shifts in the underlying security state, with transition probabilities that can be linked to observable indicators.
- Covariate-driven intensity models allow diplomatic, military and economic indicators to shift the rate, at the cost of requiring those indicators to be available in real time.
- Hierarchical Bayesian structures allow information to be shared across zones with limited individual data while preserving zone-specific differences.
Added structure buys realism at the cost of parameters. In sparse-data settings, a model with many parameters can fit the observation window closely and generalise poorly. The discipline is to add structure only where it is supported by evidence and where the added realism changes a decision.
Validation when prediction is not the objective
A geopolitical model that fails to predict a specific attack has not failed. It would fail if the frequency of events it deems unlikely turned out to be systematically higher than the model implies, or if its intensity estimates responded to information in the wrong direction.
- Calibration testing: over many periods, do events occur at approximately the frequency the model assigned?
- Residual analysis for point processes: after time-rescaling, are the transformed inter-event times consistent with the assumed process?
- Out-of-time testing: fit on one window, evaluate on a later window the model did not see.
- Benchmarking against a simple baseline, such as a constant rate or a recent-window empirical rate; complexity must earn its place.
- Stability testing: how much do parameters move when the estimation window shifts by a few months?
Underwriting applications
- Technical rate ranges by zone and vessel profile, refreshed on a defined cycle and on trigger events.
- Accumulation monitoring against live in-force transits, with pre-set action thresholds.
- Early-warning indicators that link a measured intensity change to a defined underwriting response.
- Scenario suites for capital assessment and reinsurance structuring.
- Consistent documentation for internal challenge, audit and supervisory discussion.
Practical example
A model assigns a baseline intensity to a zone during a quiet regime and a substantially higher intensity in an active regime, with an estimated probability of transition between them each week. Following an incident, a self-exciting component temporarily raises intensity, decaying over a few weeks.
In practice this produces three underwriting behaviours that a static rate would not: an immediate, quantified rate increase after an event; a gradual decay rather than an abrupt return to baseline; and a step change if the regime indicator flips. Each is defensible to a client because the mechanism can be explained, and each is auditable because the inputs are recorded.
Limitations and caveats
- Event data under-reports incidents in areas with limited observation, biasing intensity estimates downward where risk may be highest.
- Political change is driven by decisions that leave no statistical signature until they occur.
- Parameter estimates in sparse-data zones carry wide uncertainty that should be reported alongside point estimates.
- Models calibrated during one conflict may transfer poorly to another with different actors, tactics and targets.
- Output supports professional judgement and is not a forecast of specific events.
Conclusion
Quantitative geopolitical models earn their place by making underwriting consistent, explainable and scalable — not by anticipating events.
Their honest presentation includes wide uncertainty bands, named data limitations and an explicit statement that qualitative assessment remains part of the decision.
References
- Armed Conflict Location & Event Data Project methodology — ACLED
- UCDP Georeferenced Event Dataset codebook — Uppsala Conflict Data Program, Uppsala University
- Review of Maritime Transport — United Nations Conference on Trade and Development
Author bio
Jonas Mohamed Osman Abdelghafour, known as Yonas Osman is an actuary, FRM and financial risk professional specialising in banking, insurance, model risk, capital modelling and quantitative risk management.