3.4pp
MEDIAN PREDICTION ERROR
Half of all predictions land within 3.4 percentage points of the actual crash. Tested across 67 token-scenario pairs in 4 crises. SOL during FTX: predicted -67.6%, actual -67.9%. AAVE during LUNA: predicted -52.8%, actual -51.7%.
82%
PREDICTIONS WITHIN 10pp
55 of 67 predictions land within 10pp of actual crash magnitude. For the FTX collapse, average error was just 3.6pp across 16 tokens. BTC during Oct 2025 flash crash: predicted exactly -13.0%.
5
DIFFERENT CRISIS TYPES
Exchange failure (FTX), stablecoin death spiral (LUNA), lending contagion (3AC), and infrastructure flash crash (Oct 2025). Per-crisis-type calibration: different mechanisms propagate risk differently, and the model accounts for this.
What this means: When you run a stress test below, the predicted losses are calibrated against real crisis data with a median error of 3.4 percentage points. The model recognizes different crisis mechanisms — — exchange failures, DeFi collapses, and flash crashes each propagate risk differently. SOL predicted -67.6% vs actual -67.9% in the FTX collapse. 82% of all predictions land within 10 percentage points of the actual loss.
Full methodology →