How Convex Works

Four composite indicators, CVRP, CNLI, CRAI, and NVI, each synthesizing multiple proven signals into a single actionable reading. Covers recession probability, net liquidity, cross-asset risk appetite, and narrative velocity.

• , CVRP: 5-channel recession probability (yield curve, Sahm Rule, claims, credit, LEI)
• , CNLI: Fed balance sheet minus RRP minus TGA, actual market liquidity
• , CRAI: 5 cross-asset ETF ratio z-scores spanning equities, credit, sectors, EM, banks
• , NVI: Real-time narrative acceleration from 46 editorially diverse news sources

Over 300 live data sources ingested, validated, and transformed into a unified macro data estate. External feeds from FRED, GDELT, CFTC, Polymarket, and dozens of RSS sources, plus internally derived composite signals.

• , 300+ FRED economic series across rates, inflation, credit, liquidity, labor, activity
• , GDELT geopolitical event stream with 80+ trusted domain whitelist
• , CFTC Commitments of Traders: speculative positioning + percentile ranks
• , Polymarket prediction markets: real-time market-implied probabilities
• , Dozens of RSS feeds across editorial tiers (official government, think tanks, quality news, financial wires, aggregators) for narrative-velocity signal

A Dynamic Factor Model reduces the macro series universe to latent factors via UD Kalman filtering and EM estimation. Hamilton 2-state regime detection classifies normal vs. stress environments. A bilateral stress tensor quantifies real-time geopolitical risk from event data.

• , Dynamic Factor Model: PCA initialisation → multi-start EM → daily Kalman filter
• , Hamilton regime filter: Markov-switching 2-state model on factor scores
• , Bilateral stress tensor: CAMEO channel scoring across military, economic, diplomatic
• , Eigenvalue crisis monitor: tracks variance compression across principal components
• , Transfer entropy: directed information flow between macro series

The part nobody else has. Scenarios update every 6 hours against live data using Bayesian math, not gut feel. The system discovers new scenarios autonomously, validates them through multi-source evidence gating and causal chain tracing, retires dead ones automatically, and calibrates its own accuracy over time. An LLM and a Bayesian engine run in parallel and check each other.

• , Continuous 6-hour Bayesian probability updates, not static PDFs
• , Autonomous scenario discovery across 6 independent data sources
• , Causal cascade tracing: oil spike → news clustering → positioning shifts → inflation expectations
• , Dual-brain architecture: creative LLM + disciplined Bayesian engine, divergence surfaced as signal
• , 22-gate article pipeline: enrichment → evidence → probability → generation → validation

Hand-curated long-form analysis follows a verified-facts protocol. Every numerical claim in research-driven content is sourced to either a fact ledger (each entry has a verification URL) or a hand-reviewed comparison-pairs narrative. The catch test: for any number on the page, the writer must produce a URL or code reference. Estimates and "plausible" figures are not allowed.

• , Verified-facts ledger: ~130 sourced data points across April 2026 snapshots and historical anchors from 1973 forward
• , Cross-source verification: single-source claims are not enough; institutional/auditor-grade rigor required
• , Self-audit before push: every numerical claim re-checked against the ledger; fabrications caught and corrected pre-deploy
• , Editorial dual-attribution: AI-written by Convex Research Desk, human-edited by Ben Bleier (visible byline + JSON-LD editor field)
• , Live snapshot dating: every fact carries an as-of date so readers and AI assistants can judge freshness
• , Used by: 41 hand-curated scenario × asset crown-jewel pages at /what-happens-when/{slug}/{asset}