Methodology

1. Architecture Overview

ZARQ Crypto Risk Intelligence consists of two complementary models that serve different purposes and use different data.

These are not the same model with different names. They measure different things, use different data sources, and carry different weights. They work together in pair selection: Trust Score drives relative ranking, DtD filters distress risk. Together with the Structural Collapse filter, they form the basis for ZARQ's early warning system — which detected 113 out of 113 token deaths with 98% precision in out-of-sample testing.

2. Trust Score

Trust Score assesses the overall reliability and quality of crypto entities (tokens, exchanges, DeFi protocols) based on publicly available data. Every token is scored 0–100 across five pillars and mapped to a Moody's-style letter scale from Aaa to D.

Five Pillars

Rating Scale

Trust Scores are mapped to a Moody's-style letter scale used throughout the platform:

Entity-Specific Scoring

Trust Score has separate scoring functions for tokens, exchanges, and DeFi protocols. The pillars and weights are identical, but input variables are adapted. Tokens use market cap, ATH recovery, contract verification, and GitHub metrics. Exchanges use CoinGecko trust score, proof of reserves, and trading volume. DeFi protocols use TVL, audit status, hack history, and chain deployments.

Validation

Retrospective analysis shows consistent separation: collapsed exchanges averaged 5.0/100 (vs. 44.5 platform average), collapsed tokens averaged 30.9/100. All major 2022–2023 collapses scored below platform average.

3. Distance-to-Default (DtD)

Adapted from Merton's structural credit model used by credit agencies for bonds, DtD measures how many standard deviations a token is from critical failure. Scale 0–5, where 5 = fully healthy and 0 = imminent collapse risk. Below 2.0 is the danger zone. Below 1.0 is imminent risk.

Seven Signals

Override Rules

Alert Levels

DtD Trend Classification

4. Structural Collapse & Stress Detection

ZARQ's primary early warning system. Rather than relying on ML crash prediction models, the structural weakness filter uses a rule-based approach that has been validated out-of-sample with exceptional results.

Four Weakness Signals

Alert Levels

Structural weakness ≥ 2 triggers STRUCTURAL STRESS — fundamentals weakening, two weakness signals detected, requires monitoring. When a third signal activates, the token escalates to Structural Collapse.

Structural weakness ≥ 3 triggers STRUCTURAL COLLAPSE — this token is breaking apart. Historically, 98% of tokens with this profile lost more than half their value. Most never came back.

Out-of-Sample Validation

Period: January 2024 – February 2026. Universe: 207 tokens.

5. Crash Probability

90-day forward probability combining DtD trajectory, TVL divergence, contagion exposure, and historical default patterns. Calibrated on 393 crash cycles. Lookup: (trend, alert_level) → P(crash >30% within 90 days).

6. Contagion Model

The contagion model maps dependency relationships between tokens through shared liquidity pools, ecosystem membership, bridge exposure, oracle dependencies, stablecoin reliance, and exchange concentration. Each token receives a Contagion Score (0–10) representing vulnerability to cascading failures.

Dependency Channels

Scenario Analysis

Three pre-built scenarios (major exchange collapse, stablecoin depeg, ecosystem contagion) plus custom portfolio input. Retroactive case studies validate the model against FTX, LUNA/UST, and 3AC — all using the same calibrated betas applied to current data.

Interactive contagion map with scenario analysis available at /contagion.

7. Portable Alpha Strategy

The Portable Alpha strategy combines Trust Score-based pair selection with DtD distress filtering and bear market detection. Conviction-ranked long/short pairs generated from rating spreads and DtD differentials within investment-grade tokens.

Architecture

1. Trust Score-based pair selection — long top-quartile, short bottom-quartile within the same rating class (IG_MID)
2. DtD-based distress filtering — do not short tokens with DtD < 1.5
3. Bear market detection — skip when BTC monthly return < −15%
4. Capital allocation — allocate between pairs portfolio and cash/BTC based on regime

Implementation

Universe of ~85 major tokens, excluding stablecoins. Rating class: IG_MID only (best-performing class for pair alpha). Top 5 pairs per month based on conviction score (40% spread + 60% DtD differential). Max 2 pairs per token. 90-day holding period. Return cap: ±100% per leg.

Three Portfolios

Paper trading launched March 1, 2026 with SHA-256 hash-chained audit trail. Every signal logged before market open. No hindsight. No adjustments. Live tracking at /paper-trading. Backtest results and historical performance at /track-record.

8. API Reference

Every model output maps to an API endpoint with defined JSON schema. All endpoints available at zarq.ai/v1/crypto/. Free during beta — 1,000 calls/day, no API key required.

Example: Safety Check

GET /v1/crypto/safety/bitcoin

{
  "data": {
    "token_id": "bitcoin",
    "safe": true,
    "risk_level": "LOW",
    "trust_grade": "A2",
    "dtd": 3.03,
    "alert_level": "SAFE",
    "crash_probability": 0.03,
    "flags": []
  },
  "meta": { "version": "v1", "response_ms": 12 }
}

9. Machine-First Architecture

ZARQ is built API-first. The dashboard is a showcase — the API is the product. Every data point visible on the website is available through a documented JSON endpoint.

AI Discoverability

Optimized for autonomous AI agent consumption: llms.txt and llms-full.txt provide structured context for language models. robots.txt includes AI bot directives. Schema.org markup on all pages. AI-citable summary comments embedded in HTML source. MCP (Model Context Protocol) server available at zarq.ai/mcp/sse.

Bulk Data

All 198 token ratings available as compressed JSONL at /data/crypto-ratings.jsonl.gz, licensed under CC BY 4.0. Includes Trust Score, DtD, pillar scores, and structural weakness signals. Metadata header with citation format.

Infrastructure

Self-hosted on dedicated hardware. SQLite for crypto data (zero-overhead reads), PostgreSQL for agent data. Cloudflare Tunnel for TLS termination. Automated daily pipeline: data collection → scoring → signal generation → NAV calculation → audit logging. System health monitored every 5 minutes with automatic self-healing.

10. Case Studies

Retroactive analysis of major crypto collapses using the Trust Score methodology. These scores are computed on current data, not pre-collapse data — they demonstrate that the scoring methodology captures structural risk factors.

FTX (November 2022)

Trust Score: 5.0/100 (Grade F). No proof of reserves, opaque corporate structure (Bahamas + Antigua), FTT token used as circular collateral, Alameda Research conflicts of interest. Losses: $8.0B.

Terra/LUNA (May 2022)

Trust Score: 57.2/100 (C+) for LUNA, 29.9/100 (D) for UST. Algorithmic stablecoin with no real reserves, unsustainable 20% APY via Anchor Protocol, known death-spiral risk mechanism. Combined losses: $58B.

Celsius Network (June 2022)

Trust Score: 5.0/100 (F). Below-platform scoring across all five pillars. Losses: $4.7B.

Separation Analysis

11. Limitations

Data History

The crypto market has approximately 10 years of history compared to traditional finance's 100+. All statistical results should be interpreted with this constraint.

What The Models Do Not Capture

Hacks and technical exploits, regulatory surprises, insider fraud (FTX-type events), political events, key personnel departures, and smart contract bugs. These are inherently unpredictable from on-chain and market data alone.

Execution Costs

Pairs backtest does not include transaction costs. Estimated cost: 0.5% per trade (entry + exit), approximately 1% round-trip per pair.

Sample Sizes

Crash probability table is based on 393 crash cycles with per-cell counts not reported. Structural Collapse OOS validation covers 207 tokens over 26 months.

Concentration Risk

The Portable Alpha strategy runs 5 pairs per month from ~85 tokens. This represents high concentration in individual positions. Portfolio-level drawdown depends on allocation variant chosen.