Projects  /  Featured

QGI

Quantitative Geopolitical Intelligence: a solo-built platform that reads six decades of national indicators to find countries tracing the trajectories that preceded past crises.

An applied-machine-learning platform that reads six decades of national indicators to find countries tracing the same trajectories that preceded past crises, and surfaces the real historical precedents behind every signal.

Live at qgintelligence.com ↗ Presented a QGI workshop · 8th PPE Conference, University of Witten/Herdecke Solo-built · 7 repos · Multi-account AWS

Status: research prototype; models and recipes under academic review. QGI reports how strongly a country resembles historical precedents, ranked, evidenced, with the uncertainty stated, rather than issuing probabilistic forecasts.

QGI is a solo-built, seven-repository system running on a multi-account AWS backbone, live at qgintelligence.com. Its current methodology is the Trajectory Engine: a fleet of per-event-type models, each learning which indicators matter, over what horizon, and how closely a country today resembles the historical cases behind that event type. A separate statistical scoring layer, older than the fleet and unrelated to the retired recipes, still contributes country scores alongside it. The platform's original engine, the Classic Cascade, a pure-statistics correlation pipeline that mined cross-country indicator correlations into event recipes, is retired: it no longer feeds the live pipeline, and is described below for its own engineering history rather than as part of the current system.

Methodology

Two generations, one platform

QGI has been built twice. Pick a generation to see how it was engineered; only the Trajectory Engine is live.

Trajectory Engine: the ML layer

Each recipe becomes its own model, architecture chosen by how much positive-class data it has: a regularised logistic regression when positives are thin, XGBoost once there are enough. A univariate pre-filter narrows the indicator catalogue to the handful that carry signal for that event type, then a greedy search selects the final set and the horizon over which the pattern plays out, cross-validated throughout, leave-one-country-out, with bias-corrected bootstrap confidence intervals, so a model is never scored on a country it trained on. Across the fleet, that validation discipline is the point, not a single number: median out-of-fold AUC is 0.633 (range 0.535–0.741), a modest but real early-warning signal. Surviving models rank today's countries by resemblance to the historical precedents, never a probability.

42Per-Event Models
71Event Categories
250Indicators · 8 Providers
LR + XGBoostPer-Event Architecture
Low-CostSpot Fleet

The current modelling generation: a fleet of per-event-type models on the current indicator catalogue. A separate statistical scoring layer, older than the fleet, still contributes country scores alongside it.

  1. 1 · Ingest

    • Amazon S3
    • Parquet
    • 250 indicators · 8 providers
    • World Bank · V-Dem · UCDP · FAOSTAT
    • UNHCR · ILOSTAT · OWID · UNDP
    • over 10,000 verified events
    • 71 categories · 1903–2026

  2. 2 · Feature store

    • Amazon S3
    • AWS Glue
    • Athena
    • Columnar feature store
    • Glue-catalogued, queried in SQL
    • Curated event labels

  3. 3 · Orchestration

    • Step Functions
    • AWS Batch
    • DynamoDB
    • Map state, one run per recipe
    • Graviton Spot · c7g / m7g
    • Python 3.13 · arm64
    • Coordination counters

  4. 4 · Model fleet

    • AWS Batch
    • Amazon S3
    • XGBoost
    • scikit-learn
    • 42 models, one per event type
    • logistic regression or XGBoost
    • per-recipe models written to S3

    Gates

    • coverage gate
    • chance gate
    • 7-confound battery

  5. 5 · Publish

    • Amazon S3
    • Vercel
    • Next.js
    • baked static JSON
    • qgintelligence.com

    Also scoring

    • a separate statistical layer
    • older than the fleet, still contributes country scores

Multi-account AWS org · eu-central-1 · Terraform per-environment state · GitHub Actions OIDC, no long-lived credentials

  1. Recipe + Labels
  2. 1

    Feature Substrate

    1. Amazon S3 · Parquet

      Feature Substrate

      250 macro indicators across 8 providers (World Bank, V-Dem, UCDP, FAOSTAT, UNHCR, ILOSTAT, OWID, UNDP) as a shared columnar feature store, plus curated and verified event labels (positives).

  3. indicators + labels
  4. 2

    Per-Recipe Walk F-ML-1

    1. Python · scikit-learn

      Univariate Pre-filter

      Ranks 250 indicators by leave-one-country-out signal and keeps the top 40, capping the expensive search so the indicator pool can grow without exploding compute.

    2. Greedy Search

      Indicator Selection + Window Search

      Greedy forward selection over indicators, jointly searching the canonical horizon {3 … 25 years}; a candidate joins the recipe only if it lifts held-out AUC.

    3. XGBoost / Logistic

      Train + Leave-One-Country-Out AUC

      XGBoost per recipe (logistic fallback for sparse classes), scored by leave-one-country-out cross-validation so no country is ever evaluated on itself.

    4. Clears the LOCO-AUC floor?

      clears floor below floor
  5. canonical shape + indicator weights
  6. 3

    Similarity-as-Risk Scoring

    1. SHAP (Shapley additive explanations)

      Similarity-as-Risk Scoring

      Each country's recent trajectory is compared with the recipe's precedent pattern and reported as a 0–100 resemblance score with a confidence band.

  7. 0–100 resemblance + precedents
  8. 4

    Recipe-Fleet · AWS Batch

    1. Step Functions · Batch · Spot

      Parallel Fleet Fan-Out

      A canary-gated Step Functions Map fans every runnable recipe across right-sized 2-vCPU Graviton Spot instances (c7g / m7g), writing per-recipe models back to S3, a full fleet run costs a few dollars.

  9. Trend Tracker · live

Resemblance ranks, not probabilities  ·  leak-guarded (leave-one-country-out)  ·  research prototype & under academic review

Validation & Integrity

Every model in the fleet is tested the same way: leave-one-country-out cross-validation, so a model is never scored on a country it trained on, with bias-corrected bootstrap confidence intervals, time-matched negative sampling, and three-cohort stratification. Before a recipe is even attempted it has to clear a coverage gate (enough positives, enough negatives, enough surviving positives across enough countries) and a chance gate (its bootstrap confidence interval has to clear 0.50), then it is stress-tested against a battery of seven confounds designed to catch the platform's own false positives. Pre-registered falsification criteria and a forbidden-actions list were written before any model was trained.

An early result reported an AUC of 0.9147, computed on 10 of 174 positive cases. It was wrong: recomputed on the full set, it corrected to about 0.4967, no better than random guessing. The claim was retracted publicly, and the one recipe that had briefly graduated was reset. A fleet-wide audit then re-measured all 42 models under the same protocol; median out-of-fold AUC is 0.633, described above.

Platform & Stack

Python XGBoost scikit-learn SHAP (Shapley additive explanations) LOCO Cross-Validation Pearson Correlation AWS Step Functions AWS Lambda AWS Batch Graviton Spot AWS Glue AWS Athena Amazon S3 DynamoDB Terraform GitHub Actions OIDC GDELT · BigQuery Next.js Parquet Docker