Forecasting Model Operations for Conflict Analytics

Forecasting conflict outcomes—whether territorial, political, or escalatory—is among the most challenging analytical tasks in geopolitics. The Ukraine war has become a major proving ground for forecasting methodologies, attracting prediction markets, expert elicitation programs, quantitative model developers, and comparative political scientists competing to produce calibrated probability estimates on questions ranging from near-term battlefield developments to long-term war termination timelines. This article reviews the primary forecasting methodologies applied to the Ukraine conflict and assesses their comparative track records.

Prediction Markets and Forecasting Platforms

Prediction markets aggregate distributed information through financial incentives—forecasters stake real or virtual currency on specific outcomes, producing market prices interpretable as probability estimates. Metaculus, Manifold Markets, Polymarket, and Good Judgment Open host hundreds of Ukraine-related questions covering battlefield developments, political outcomes, ceasefire probabilities, and military escalation. Metaculus, as of early 2026, hosts over 800 resolved Ukraine war questions. Market performance on Ukraine has been mixed: prediction platforms correctly estimated high probability of the full-scale invasion ahead of 24 February 2022 (Metaculus community was at ~50% in January 2022 when intelligence was building), but substantially underestimated Ukrainian battlefield resilience in the conflict's early weeks. Polymarket's real-money architecture provides stronger incentive alignment but concentrates participants willing to risk capital, potentially skewing toward overconfident traders rather than careful epistemic forecasters.

Expert Elicitation Methodologies

Expert elicitation structures judgments from domain-specialist panels using formal methods to overcome cognitive biases documented in unstructured expert opinion. The RAND Delphi method and Tetlock's Structured Analytic Techniques provide the primary frameworks. Applied to the Ukraine conflict, expert panels organized by RAND, IISS, and the European Council on Foreign Relations have been asked to estimate probability distributions over questions like "probability of Russian forces capturing Kharkiv city within 12 months" or "probability of ceasefire before January 2025." Key methodological features: multiple rounds of elicitation with feedback to reduce anchoring; incentive scoring (experts receive accuracy feedback to promote calibration); independence maintenance (preventing premature convergence on a dominant opinion); and diversity management (ensuring panel includes regional specialists, conflict analysts, and quantitative modelers whose different expertise reduces correlations between errors).

Forecasting Platform Track Record Comparison

Quantitative Ensemble Models

Model ensembling—combining multiple individual forecasting models into a weighted average—consistently outperforms single-model forecasts across domains, including conflict analytics. In the Ukraine context, quantitative ensemble models draw on historical conflict data (UCDP Armed Conflict Dataset, PRIO Battle Deaths Dataset), current OSINT feeds, and structural variables (force ratios, logistics capacity, air superiority metrics) to produce probabilistic territorial outcome estimates. The Uppsala University VIEWS (Violence and Impacts Early-Warning System) project generates conflict risk forecasts for Ukraine using machine learning ensemble methods trained on historical African and Middle Eastern conflicts, recalibrated for European context. Forecasting ensemble performance generally beats individual analyst predictions in well-defined, short-horizon questions (30-90 days) but degrades at longer time horizons where structural uncertainty dominates stochastic model assumptions.

Challenges Specific to Ukraine Forecasting

Several factors make Ukraine forecasting distinctly difficult. Political decision uncertainty—whether the US Congress will pass aid packages, whether Germany will approve Taurus missiles—introduces non-military pivots that quantitative models miss. Russian strategic unpredictability (escalation risk calibration) involves decisions made by a small group with internal deliberative processes opaque to outside forecasters. Ukrainian operational creativity has repeatedly surprised analysts on both sides, as non-linear outcomes (Kharkiv counteroffensive, Kherson liberation) followed periods of apparent stalemate. Reference class uncertainty is also severe: the last large-scale European inter-state war was fought with pre-nuclear, pre-precision, pre-drone technology—historical base rates for modern conflict dynamics are thin.

Calibration and Scoring

Forecast quality is assessed through calibration (probability estimates reflecting true event frequencies) and resolution (probabilities that actually discriminate—not clustered at 50%). Brier scores and log scores measure combined calibration and resolution. Superforecasters identified through the Good Judgment Project achieve Brier scores roughly half those of conventional expert panels, largely through disciplined probability updating rather than superior domain expertise. Ukraine-focused analysis should target explicit calibration maintenance: analysts who issued "90% confident" assessments that proved wrong must update their uncertainty priors. Institutional resistance to this accountability is a known failure mode in government and think-tank analytical cultures.

FAQ

What is a superforecaster and are they relevant to Ukraine analysis?

Superforecasters, identified in Philip Tetlock's extensive research program, are individuals who demonstrate consistently superior probabilistic prediction across domains through disciplined belief updating, active open-mindedness, and rigorous self-scoring. The Good Judgment Project has applied superforecaster panels to Ukraine questions with somewhat better calibration than expert panels, though the advantage is smaller on high-uncertainty geopolitical questions than on near-term factual questions where more information is available.

Why did most forecasters underestimate Ukrainian resilience in early 2022?

Most early-2022 forecasts used historical base rates from conflicts between asymmetric powers (Soviet-Afghan, US-Iraq) where the conventional military advantage strongly predicted rapid victory. Ukrainian institutional resilience, Western aid speed, and Russian operational failures—particularly logistics and combined arms coordination—were underweighted because recent historical analogies were poor guides to this specific war's dynamics. The lesson: structural model inputs must reflect the specific combatants' capabilities, not only historical averages.

How does Metaculus handle resolution of contested outcomes?

Metaculus has formal resolution criteria written into each question at creation, specifying which sources will be used for resolution and under what conditions. For contested outcomes (territorial control in active battle zones), questions typically resolve based on internationally recognized sources or majority-community assessment. Unresolvable or ambiguously worded questions are annulled (bets refunded), providing an incentive for careful question drafting that builds in clear resolution criteria.

Can AI language models be used for conflict forecasting?

LLMs (GPT-4, Claude, Gemini) show some capability in short-horizon conflict forecasting when provided with relevant OSINT context, primarily by synthesizing large-context information faster than human analysts. However, they lack calibrated uncertainty quantification and are susceptible to recency bias and training data cutoffs. Current best practice uses LLMs as information synthesis aids for human forecasters rather than as autonomous forecasting systems.

What is the VIEWS project and how does it forecast for Ukraine?

VIEWS (Violence and Impacts Early-Warning System) is a Uppsala University/Peace Research Institute Oslo collaboration that generates probabilistic fatality and conflict escalation forecasts using ensemble machine learning models trained on historical UCDP conflict data. For Ukraine, VIEWS provides rolling 6-month probabilistic forecasts of conflict fatality levels by administrative division, calibrated against its historical training data and updated monthly with new UCDP-coded events. It demonstrates good calibration for sustained conflict duration in established conflict zones but lower accuracy for sudden outbreak/de-escalation transitions.

Sources

1. Tetlock, P. and Gardner, D., Superforecasting: The Art and Science of Prediction, Crown, 2015.
2. Metaculus, Ukraine Conflict Question Track Record, metaculus.com/questions/?topic=ukraine, 2026.
3. VIEWS Project, Conflict Forecasting Methodology and Ukraine Results, Uppsala University, 2024.
4. Sundberg et al., Introducing the UCDP Georeferenced Event Dataset, Journal of Peace Research, 2013.
5. Good Judgment Project, Superforecaster Panel on Ukraine Conflict, project report, 2024.