Liquidation Cascade Risk

Liquidation Cascade Risk

When This Skill Activates

Apply this skill whenever any of the following pre-cascade signals are present on HyperLiquid perpetual markets:

• Open interest (OI) has increased > 15% over the prior 24h while price is range-bound or declining
• OI-weighted average leverage on the asset is estimated > 10× (infer from OI / market cap ratio when direct leverage data is unavailable)
• Price approaches a known liquidation cluster level (derived from OI distribution or historical wicks)
• Bid-side depth within 0.5% of mid has dropped > 30% versus the prior 1h average (thin book = amplified cascade velocity)
• Funding rate has been > 0.05%/h for ≥ 6 consecutive hours (extreme carry = crowded long leverage)
• A large liquidation event (> $5M notional on any single asset) has already fired — first liquidation often triggers cascade
• CVD (cumulative volume delta) diverging strongly negative vs. price (hidden sell pressure before a long-side cascade)

Core Principle

Liquidation cascades are reflexive, self-reinforcing events: forced long closures drive price down, which liquidates more longs, which drives price further down. On thin-book perpetual markets like HyperLiquid, a cascade can clear 3–8% of price in under 60 seconds with near-zero mean-reversion in the first 30 seconds.

Rule: Never hold a leveraged long into a confirmed cascade. Size aggressively to exit at the first liquidation signal, not after confirmation. Being wrong costs slippage; being right saves the position.

Signal Classification and Cascade Probability

Pre-Cascade Regime Indicators

Probabilities are additive uplift estimates, not independent probabilities. When 3+ signals are simultaneously active, treat cascade risk as HIGH and apply the full defensive protocol below.

Cascade Severity Classification

Step-by-Step Pre-Cascade Protocol

1. Fetch OI and Liquidation Data

GET /info → meta + assetCtxs
  → assetCtxs[i].openInterest       (current OI in base asset)
  → assetCtxs[i].funding            (current predicted 1h rate)
  → assetCtxs[i].markPrice          (mark price for cluster proximity)

GET /info → liquidations (via WebSocket: subscribe to l2Book + trades)
  → monitor for large single-fill liquidation events

Compute OI change:

oi_change_pct = (oi_now - oi_24h_ago) / oi_24h_ago * 100

2. Assess Book Depth Fragility

GET /info → l2Book(coin)
  → sum bid sizes within 0.5% below mid  = bid_depth_near
  → compare to 1h rolling average         = bid_depth_avg_1h

fragility_score = max(0, 1 - (bid_depth_near / bid_depth_avg_1h))
# fragility_score > 0.30 = dangerously thin book

3. Score Active Signals and Classify Risk

def cascade_risk_level(signals: dict) -> str:
    score = 0
    if signals['oi_change_24h_pct'] > 30:   score += 3
    elif signals['oi_change_24h_pct'] > 15: score += 2
    if signals['funding_elevated_hours'] >= 6: score += 2
    if signals['book_fragility'] > 0.30:    score += 3
    if signals['near_oi_cluster']:          score += 3
    if signals['large_liq_fired']:          score += 4
    if signals['cvd_diverging']:            score += 2

    if score >= 8:  return 'CRITICAL'   # exit all longs immediately
    if score >= 5:  return 'HIGH'       # reduce longs 50–75%
    if score >= 3:  return 'ELEVATED'   # reduce longs 25–40%
    return 'NORMAL'

4. Execute Defensive Action by Risk Level

CRITICAL override: When large_liq_fired signal fires (+4 score), the CRITICAL threshold can be hit with only one other signal active. Do not wait for full scoring — execute exit the moment a confirmed large liquidation event is observed AND any one other signal is active.

During a Cascade — Live Event Protocol

Once a cascade is confirmed (price dropping > 1% in < 2 min with rising liquidation volume):

Do NOT

• Add to long positions to "average down" — cascades have no floor until OI is cleared; the market will liquidate your averaging attempt too
• Place resting limit bids in the cascade path — bids will be hit immediately, not providing price improvement but capturing falling knives
• Use stop-loss orders inside the cascade — stop-limit orders may not fill; use market orders for exit during a cascade
• Attempt to short the cascade if not already positioned — entry during a cascade is extremely high slippage; the best edge is pre-positioning

Do

• Execute all long exits via market order (not limit) — slippage cost is always lower than the cost of not exiting during a major cascade
• Monitor the liquidation feed for cascade exhaustion signals:
  • Liquidation volume decelerating (fewer liquidations per 30s)
  • Bid depth beginning to rebuild in the l2Book
  • Funding rate dropping sharply (extreme shorts beginning to dominate)
  • CVD turning positive (buyers absorbing supply)
• Log the cascade start price, trough price, and recovery level for the evolution curriculum agent's liquidation-cascade task family

Post-Cascade Re-Entry Protocol

Cascades create re-entry opportunities once OI has been flushed and liquidation pressure exhausts. The post-cascade long is one of the highest-probability setups in perpetual markets.

Re-Entry Criteria (all must be met)

1. OI has declined ≥ 15% from pre-cascade peak (deleveraging confirmed)
2. Funding has normalised to < 0.02%/h (crowded longs cleared)
3. Bid depth has recovered to ≥ 80% of pre-cascade 1h average
4. Price has stabilised — no new lows for ≥ 5 consecutive 1m candles
5. CVD is turning positive — net buying pressure visible on tape

Re-Entry Sizing

Post-cascade re-entries use reduced initial size due to residual uncertainty about whether the cascade is fully exhausted:

re_entry_size = base_size * 0.5   # initial leg

# Add second leg when:
#   price holds above entry for > 15 min AND
#   OI begins rising again (new longs building = recovery confirmed)
re_entry_size_full = base_size * 1.0

Stop-loss: place below the cascade trough, not below entry — the trough is the structural low; a breach invalidates the recovery thesis.

Failure Modes to Avoid

• Treating every OI spike as a cascade warning: OI can rise sustainably in genuine trend regimes. Require ≥ 2 signals before elevating risk level.
• Exiting too early on minor cascades: Score < 5 with no large liq fired does not warrant full exit. Over-trading defensive exits erodes alpha.
• Re-entering before OI flush is confirmed: Premature re-entry during a "dead-cat bounce" inside a larger cascade is the single most common post-cascade loss. OI must be declining before re-entry.
• Ignoring cross-asset contagion: BTC cascade → ETH cascade within 1–3 minutes is the historical norm. Apply CRITICAL protocol to all correlated assets when BTC cascade fires, not just BTC.
• Using cascade as a routine short entry: Cascades are violent; short entries mid-cascade have extreme slippage and risk sharp V-shape reversals. The short opportunity was in pre-positioning, not during the event.
• Forgetting the funding kill-switch interaction: The high-funding-carry-avoidance skill's kill-switch (funding > 0.15% AND drawdown > 1.5%) should already have reduced exposure before a cascade fires. If it has not, treat any cascade signal as an immediate override regardless of score.

Integration with Other Skills and Controls

Skill Dependencies

• high-funding-carry-avoidance: A prerequisite check. If funding carry avoidance has not already reduced long exposure in a high-funding regime, liquidation cascade risk is structurally higher. Run funding check first; cascade check is the second defensive layer.
• kelly-position-sizing-perps (risk/): Post-cascade re-entry sizing must still pass through Kelly/risk-budget constraints. The 0.5× initial re-entry is a cascade-specific overlay on top of base sizing.

Risk Engine Integration

{
  "event": "cascade_risk_assessment",
  "asset": "BTC",
  "timestamp_utc": "2026-04-07T22:30:00Z",
  "oi_change_24h_pct": 22.4,
  "funding_1h": 0.061,
  "funding_elevated_hours": 7,
  "book_fragility": 0.41,
  "near_oi_cluster": true,
  "large_liq_fired": false,
  "cvd_diverging": true,
  "cascade_score": 7,
  "risk_level": "HIGH",
  "action_taken": "reduced_long_60pct",
  "position_before_usd": 15000,
  "position_after_usd": 6000
}

Log this record to logs/metaclaw/records/ on every cascade assessment that results in a position change so it enters the RL replay buffer.

Quick Decision Tree

Is cascade_score >= 8 OR (large_liq_fired AND any other signal active)?
├── YES → CRITICAL: market-exit ALL longs immediately. Halt new longs.
│
└── NO  → Is cascade_score >= 5?
          ├── YES → HIGH: reduce longs 50–75%. Cancel resting long orders.
          │
          └── NO  → Is cascade_score >= 3?
                    ├── YES → ELEVATED: reduce longs 25–40%. Widen stops.
                    └── NO  → NORMAL: monitor. No action required.

Post-exit: monitor for cascade exhaustion.
  All 5 re-entry criteria met?
  ├── YES → Re-enter at 0.5× base size. Add second leg after 15m hold.
  └── NO  → Remain flat. Do not chase recovery.

HyperLiquid-Specific Notes

• WebSocket liquidation feed: Subscribe to trades channel and filter for "liquidation": true in the trade object. HyperLiquid does not have a dedicated liquidation stream; liquidation fills appear in the standard trades feed with a liquidation flag.
• Mark price vs. last price: Liquidations trigger at mark price (oracle-weighted), not last trade price. During thin-book cascades, last price can lag mark price by 0.5–2%. Use assetCtxs[i].markPrice for liquidation cluster proximity calculations, not mid from the l2Book.
• Cross-margin cascade amplification: HyperLiquid cross-margin accounts allow a cascade on one asset to liquidate positions on correlated assets if overall margin falls below maintenance. Monitor marginSummary.totalRawUsd and marginSummary.totalMaintMargin ratio; if ratio < 1.3, pre-emptively reduce all cross-margin longs regardless of per-asset cascade score.
• Vault LPs are cascade beneficiaries: HyperLiquid vault LPs earn from liquidation fees and benefit from OI resets. This skill applies only to directional perpetual positions, not vault LP allocations.
• Order book reset after cascade: HyperLiquid's order book fully refreshes after a cascade event. The first 30 seconds of l2Book data post-cascade is unreliable for depth assessment — wait 60 seconds before using book depth as a re-entry signal.