Krishaay Jois · Ansh Parmar · Abhyuday Sengar
An early-warning system for liquidation cascades in DeFi lending.
CascadeSignal turns a protocol's live liquidation stream into a single scalar alarm — the Hawkes branching ratio n(t) ∈ [0, 1) — that rises before a cascade and stays quiet the rest of the time. It runs on nothing but LiquidationCall events: no price oracle, no state reconstruction, no privileged data required.
Built and validated on Aave v2 and Aave v3 (Ethereum mainnet, 2020–2026), with cross-protocol transfer demonstrated on Compound v2 and Maker.
DeFi lending cascades are self-reinforcing: a price drop forces liquidations, which dump collateral, which drops prices further. By the time a cascade is visible it is already largely irreversible.
$4.11B liquidated across 49,331 events on Aave v2 alone — concentrated into a handful of crashes.
The liquidation stream is modelled as a self-exciting (Hawkes) point process. The branching ratio n(t) = α·R / (μ + α·R) measures how much of each new liquidation is triggered by prior ones. When n(t) → 1 the process is near-critical and self-feeding — the condition of a cascade.
Two design choices beyond the baseline count Hawkes:
- USD-marked blend — excitation is driven by
count + 0.5·(usd / scale), making the signal sensitive to dollar volume rather than flurries of dust liquidations. - Tolerant debounce — fire once
k = 10of the lastW = 100bars cross the threshold, so a flickering near-critical run-up fires once rather than late and repeatedly.
| Chain | Threshold | FAR (full) | FAR (out-of-sample) | Recall | Major cascades | Protectable USD |
|---|---|---|---|---|---|---|
| Aave v2 | 0.99900 | 0.38 / wk | 0.14 / wk | 98 / 98 | 3 / 3 | 81% ($3.34B / $4.11B) |
| Aave v3 | 0.99980 | 0.16 / wk | 0.18 / wk | 45 / 45 | 1 / 1 | 22% ($1.27B / $5.77B) |
- 4 / 4 major cascades caught across both chains, 100% episode recall.
- Both chains well under a 1 false alarm / week budget.
- Median lead: +6.5 min before onset, +10 min before 10% of dollar damage, +19 min before half.
The starred operating points sit below the 1/week budget with full recall on major cascades. Lead ranges from 7–65 minutes before onset on Aave v2; the fast Aave v3 episode is detected at onset (the liquidation stream has no earlier signal to fire on).
Out-of-sample years (unseen during calibration) are no worse than in-sample. Pooled OOS FAR is 0.14/wk on v2 and 0.18/wk on v3. The single above-budget year (v3 2026) coincides with genuine market-wide deleveraging — likely an undercount of true positives rather than noise.
The alarm gets ahead of 79–96% of the liquidated USD for each major cascade, and $3.34B of $4.11B (81%) in aggregate on Aave v2. These are upper bounds assuming instant intervention on every timely alarm.
USD-marking lifts recall from 0.86 → 1.00 on Aave v2 and 0.80 → 1.00 on Aave v3 at every false-alarm rate. At the 1/week budget the count model does not reach full recall; the USD-marked model does.
src/cascadesignal/ Python package
models/ Hawkes point process + liquidation-bar builder
live/ Live monitor: FastAPI app, web UI, state
labels/ Cascade labeler (ADR-001 D-A episodes)
state/, graph/ Position-state engine + contagion graph (roadmap)
ingest/, eval/, viz/ Data ingestion, evaluation, plotting
scripts/
run_cascade_labeler.py Builds the frozen ground-truth episodes
live/ bootstrap_fit · calibrate_thresholds · run_monitor
paper/ Paper figure generators (fig1–fig8)
experiments/ Frozen research experiments (E1 atlas, E3 Hawkes eval, T2 state gate)
docs/decisions/ Architecture Decision Records (ADR-001–008)
paper/ LaTeX research paper