Evaluation
Does the risk score provide genuine early warning, or does it just relabel a crisis after it's already visible? Measured directly against real data, not assumed.
Method: an elevated threshold — calm-period mean + 2 standard deviations of risk_score, computed only from genuine calm (non-crisis) history — checked against an objective, price-based depeg onset, independent of the score itself. Both the warning signal and the depeg onset require a sustained (2+ consecutive hours) crossing, not a single noisy hour — a real single-hour price spike in this data would otherwise have been mistaken for the depeg onset. Full methodology and the reasoning behind every choice: docs/FINDINGS.md.
| Metric | USDC · March 2023 | UST · May 2022 |
|---|---|---|
| Depeg magnitude | 11.9% ($0.8808) | 96.7% ($0.0334) |
| Lead time (first warning → depeg onset) 0h means already elevated at the earliest available data, not a failure to warn | 72h (3.0d) | 0h |
| Peak stress score (0–100 scale) | 81.2 | 56.3 |
| Elevated threshold used No genuine RELIABLE (calm) rows exist for UST in this dataset -- this is the pre-crisis-window average, not a validated calm baseline. | 7.18 | 7.18 (fixed, from USDC) |
| False alarms (sustained elevation outside the crisis window) | 12 | 34 |
| Stayed elevated through the crisis window | 168/168h | 160/178h |
| Trading volume vs. calm baseline Volume, not pool reserves/liquidity — reserve-balance history isn't available for the historical backfill | 9.3× ($3.4M → $31.4M) | 2.4× ($448K → $1.1M) |
| Recovery detected | Yes, 25h after window end | N/A — permanent collapse |
USDC · March 2023
UST · May 2022
Score is above threshold at the earliest available data for this pair — zero measurable lead time, because there's no genuine "before" in this dataset to measure from.