Methodology

6+ heterogeneous sources

Logicon ingests structured data from multiple independent, authoritative sources spanning different domains. Each source covers a distinct signal type — conflict events, macroeconomic indicators, governance quality, sanctions, and media intensity — ensuring no single data provider can create blind spots.

18 features across 4 domains

Raw data is transformed into a fixed-length feature vector that captures the essential dynamics relevant to each prediction question. Features are grouped into four domains to ensure coverage across all dimensions of geopolitical risk.

60% logistic regression + 40% decision stump forest

Predictions are generated by a weighted ensemble of two complementary model families. Logistic regression provides stable, interpretable baselines with well-understood uncertainty. The decision stump forest captures non-linear threshold effects — such as conflict intensity tipping points — that linear models miss. The 60/40 weighting balances robustness against expressiveness.

• Logistic regression: 14 features, L2 regularisation, fitted via IRLS
• Decision stump forest: 50 stumps, each splitting on a single feature threshold
• Ensemble weight: 0.60 logistic + 0.40 stump forest
• Output: raw probability estimate before calibration

Isotonic regression (PAV algorithm)

Raw model outputs are not well-calibrated — a predicted 0.70 may correspond to a true event rate of 0.62 or 0.78. Isotonic calibration applies the Pool Adjacent Violators (PAV) algorithm to map raw scores to empirically calibrated probabilities. This ensures that when Logicon says P = 0.35, roughly 35% of such predictions resolve positively.

• PAV: monotone non-decreasing step function fitted to historical (score, outcome) pairs
• Confidence intervals via bootstrap resampling (1000 iterations)
• Calibration quality measured by Brier score, log loss, and reliability diagrams
• Recalibrated automatically when drift is detected

7 autonomous stages

Logicon does not require manual retraining. When sufficient new outcomes accumulate or calibration drift is detected, a fully autonomous 7-stage pipeline activates. Each stage has explicit pass/fail criteria — if any stage fails, the current production model continues unchanged.

Complete reproducibility chain

Every prediction is stored alongside its complete computational provenance. This enables any prediction to be independently verified, reproduced, or challenged — a critical requirement for decision support in high-stakes environments.

• Input snapshot hash (SHA-256) — cryptographic fingerprint of all input data at prediction time
• Feature vector — full 18-dimensional vector stored as JSON for exact reproducibility
• Model version — parameter set ID linking to exact weights, thresholds, and calibration map
• Evidence chain — ranked list of contributing data points with polarity and weight
• Reasoning trace — natural language explanation of key drivers