Global Crisis Index

The Global Crisis Index is a continuous, real-time measurement of crisis intensity derived from global news feeds. It outputs a value between 0 and 100, where higher values indicate elevated crisis conditions across geopolitics and financial stability.

The index is designed to support robust financial exposure. It can be traded directly as a single position, paired against correlated assets, or overlayed into systematic strategies for hedging geopolitical risk across portfolios. For capital allocators seeking to measure and express this axis of price movement directly, the index provides a more precise instrument than traditional risk proxies.

Traditional geopolitical risk indicators lag events by hours or days. The Crisis Index operates on the same information stream that institutional actors use to reposition, with headlines reaching millions of financial professionals in real time.

The index is the output of a trained classification model applied to every headline, aggregated through an LBCGBM index function, a variance-scaled continuous process that preserves sharp regime transitions. When the world shifts from stability to crisis, the index responds in real time.

Early prototype trained on approximately one year of data (January 2025 to January 2026), validated against seven major geopolitical events. 16,487 data points, index range 42 to 84, peak during the Iran-Israel war. Methodology and limitations documented in full below.

66.2/100

Key Storylines

The chart tracks the index through seven distinct geopolitical storylines. Click any row to zoom the chart to that period.

Performance

What the chart shows

The index correctly identified all seven major geopolitical events in the dataset period. The Iran-Israel conflict and US bombing of Fordow produced the sharpest spike: +36% over 11 days, with the largest 7-day increase of +20.45. The Ukraine peace talks produced the largest sustained decline: -20%, with the largest 7-day decrease of -8.11. During the Caribbean Tensions period, gold crossed $4,500/oz as the crisis index surged. The index captures the same information driving safe-haven demand.

Limitations

One year of data is insufficient for robust validation. Seven storylines are encouraging but represent a small sample. Overfitting to any single source's editorial choices is a real risk, as the index measures what gets reported rather than the full space of geopolitical events. The model has yet to be tested against out-of-sample events or future crises.

Future improvements

Expanding training data across multiple years, multi-source ingestion, cross-validation against established indices (VIX), and formal backtesting against asset prices.

  • Largest 7-day increase+20.45 (Iran-Israel / Fordow)
  • Largest 7-day decrease-8.11 (Ukraine peace talks)
  • Storylines correctly detected7/7

How It Is Built

Data Ingestion & Filtering

The system ingests every headline from real-time global news feeds. Each data point is enriched with timestamp, referenced tickers, and raw text. Over the dataset period, this produces approximately 44,000 individual data points. Because many entries are multiline (containing multiple headlines), this corresponds to roughly 16,500 unique observations after deduplication and alignment.

Annotation

Labeling tens of thousands of headlines by hand is impractical. Instead, the annotation pipeline bootstraps from a small expert-labeled seed set:

  1. Start with 50 to 100 expert-labeled seed examples that define the crisis/non-crisis boundary with canonical cases.
  2. Run 5 to 10 AI annotator rounds using in-context learning. Each annotator sees the seed set plus canonical examples, then labels a batch of unlabeled headlines.
  3. Synthesize through consensus: where annotators agree, accept the label. Where they disagree, flag for human review.
  4. Quality control via inter-annotator agreement metrics, confidence thresholding, and systematic failure tracking across rounds.

This produces a training set larger than human annotation alone could achieve, with quality maintained through consensus rather than individual accuracy.

Model

The teacher model is a BiLSTM with learned attention, chosen for its ability to capture sequential dependencies in headline text while remaining interpretable through attention weights.

  • Embedding: vocab × 128 dimensions
  • Bidirectional LSTM: 2 layers, 256 hidden units/direction, 0.3 dropout
  • Attention: learned weighted pooling over sequence
  • Classification head: 512 → 128 → 64 → 1, Sigmoid output
  • Output: per-headline crisis probability ∈ [0, 1]

Distillation

The teacher model is compressed into a student for on-chain verification. The student uses a fixed 128-token input with mean pooling, making it deterministic and ZK-friendly with no learned attention weights. The result is approximately 50× smaller than the teacher, exported to ONNX for downstream proving.

Index Function (LBCGBM)

Per-headline crisis probabilities must be transformed into a single continuous bounded index. LBCGBM (Logit-Bounded Cumulative Geometric Brownian Motion) treats the index as a continuous stochastic process rather than a rolling average:

  1. Normalize: x = 2p − 1, centering predictions around zero.
  2. Power-law magnitudes: |x|α, α = 2.0. Strong signals are amplified, weak signals suppressed.
  3. Asymmetric volatility correction to account for class imbalance, since most headlines are non-crisis.
  4. Z-space returns with cumulative dynamics: the index has memory. Today's value depends on yesterday's.
  5. Sigmoid bounding to [0, 100].

σbase = 0.00001, α = 2.0, steepness = 1.0, initial = 50.0

Why not moving averages? Moving averages smooth regime transitions, and a sudden escalation gets diluted across the window. LBCGBM preserves sharp transitions because it models the index as a cumulative process rather than a rolling statistic.

Verification

On-chain cryptographic verification via EZKL. The student ONNX model is compiled into a ZK circuit, producing a proving key (kept private) and a verification key (deployed on-chain). For each prediction, a ZK proof attests that the published index value was correctly computed from the input headline using the published model. Consumers can verify that values were not manipulated without access to the model weights or input data.