Frequently Asked Questions

Logicon fuses four primary data sources: ACLED (Armed Conflict Location and Event Data) for real-time conflict events and fatality counts, UCDP (Uppsala Conflict Data Program) for georeferenced armed conflict data, GDELT (Global Database of Events, Tone and Language) for media event streams and sentiment, and FRED (Federal Reserve Economic Data) for macroeconomic stress indicators. All sources are normalised and aligned on a common timeline.

Raw model outputs pass through isotonic regression (Pool Adjacent Violators algorithm) which maps predicted probabilities to observed frequencies. This ensures that when Logicon states a 70% probability, approximately 70% of such predictions historically resolved as true. Calibration quality is continuously measured via Brier scores and reliability diagrams.

As forecasts resolve with real-world outcomes, the engine automatically retrains on closed predictions using walk-forward validation to prevent data leakage. Shadow models compete against the active production model and are promoted when they achieve a Brier score at least 5% lower on 30+ common resolved questions. Drift is detected via Page-Hinkley statistical tests.

Accuracy is measured via Brier scores rather than simple hit rates, because Brier scoring rewards well-calibrated probabilistic predictions. The current production model maintains a Brier score of approximately 0.055, which indicates strong calibration. All historical predictions and their outcomes are available in the audit trail for independent verification.

Logicon monitors 10 high-priority conflict regions spanning Eastern Europe, the Middle East, the Sahel, the Horn of Africa, Central Africa, South Asia, Southeast Asia, East Asia, the Caucasus, and Latin America. Each region has dedicated feature extraction pipelines assessing political, military, economic, and governance dimensions.

Each region-period observation is transformed into an 18-dimensional feature vector capturing: conflict event count and trend, fatality rates, actor proliferation, territorial control changes, GDELT media tone and event density, FRED macro indicators (inflation, sovereign stress), governance quality indices, cross-regional contagion signals, and seasonal/cyclical patterns.

Every forecast is logged with its complete feature vector, model version, calibration parameters, data sources and retrieval timestamps, and the resulting probability distribution. When a prediction resolves, the outcome is recorded alongside the original reasoning chain. This provides full traceability from raw data to final probability.

Logicon uses a weighted ensemble of 60% logistic regression (for interpretable, well-calibrated base predictions on 14 features) and 40% stump forest (50 decision stumps for capturing non-linear feature interactions). The ensemble output then passes through isotonic calibration before being presented to users.

Yes. Contact us to request an evaluation account. Evaluation access includes full forecast history, audit trails, methodology documentation, and Brier score breakdowns by region and time horizon.

Logicon generates forecasts across multiple time horizons: 30-day (tactical), 90-day (operational), and 730-day (strategic). Each horizon uses appropriately tuned feature weights and calibration curves. The self-learning pipeline maintains separate model instances for each horizon.

The system uses Page-Hinkley statistical tests to monitor model performance over time. When the Brier score drifts beyond configured thresholds (delta=0.005, alert at 0.10, critical at 0.20), automated alerts are triggered. Critical drift can escalate to automatic model retraining using the latest resolved predictions.

Shadow models are alternative model configurations that run in parallel with the production model but do not serve predictions to users. They are continuously evaluated against the active model. A shadow model is promoted to production when it demonstrates a Brier score at least 5% lower than the active model on 30 or more common resolved questions.

All incoming data passes through NaN filters, deduplication checks, and temporal alignment validation. ACLED and UCDP events are cross-referenced for consistency. GDELT streams are filtered for relevance and language quality. FRED indicators are checked for revision flags. Any anomalies are logged and flagged for review.

Logicon is designed for the NATO DIANA 2027 accelerator programme. The platform supports standard intelligence dissemination formats, provides full audit trails required for institutional accountability, and maintains data provenance records for all predictions. The architecture supports deployment in secure environments.

Logicon operates exclusively on open-source intelligence (OSINT). All data sources — ACLED, UCDP, GDELT, and FRED — are publicly available. This means the platform can be deployed without classified network requirements while still delivering operationally relevant probabilistic assessments.

Country risk scores aggregate five dimensions — political stability, economic resilience, governance quality, conflict intensity, and external pressure — into a composite risk index. Each dimension draws on multiple indicators, with weights determined by the ensemble model. Scores are updated as new data arrives and re-calibrated quarterly.

Yes. Logicon exposes a RESTful API for forecast retrieval, audit trail access, and model performance metrics. The API supports JSON output compatible with standard intelligence analysis toolchains. Custom integrations can be arranged during the evaluation phase.