Why CSI Blends Three Components (and Why Each Alone Is Insufficient)

The Cascade Stress Index (CSI) is the single number that condenses a Cascade Tensor snapshot's per-asset leverage state into something legible. We considered the obvious simpler designs and rejected each. This post explains the three-component blend and what each component catches that the others miss.

Component 1 — density at risk. The amount of notional sitting within an N% price move of its liquidation level, summed across venues and leverage buckets for the asset. Density at risk on its own is the closest single signal to 'how much would liquidate if the asset moved N%,' which is the question CSI is answering. So why isn't it the whole index? Because density at risk is invariant to where the at-risk notional sits on the topology. A million USD of at-risk longs spread evenly across six venues unwinds very differently from a million USD concentrated on one venue at one leverage bucket — the latter cascades, the former dissipates. Density at risk alone ranks the two setups identically.

Component 2 — concentration. A Herfindahl-style measure on the venue × leverage_bucket distribution of the at-risk notional, so the same total density gets a higher concentration score when it's piled on fewer cells. Concentration catches the structural difference density at risk misses — but on its own it overweights low-density-but-concentrated setups (an asset with a tiny absolute amount of leverage parked on a single venue scores high on concentration but doesn't actually pose a cascade risk). Concentration in isolation produces false alarms.

Component 3 — skew. The long-vs-short imbalance at the at-risk band, scaled to [0, 1]. A symmetric setup (longs and shorts roughly balanced at risk) unwinds in either direction without cascading — a price move triggers one side and dampens the other. A one-sided setup cascades because the price impact of the unwind reinforces the move. Skew is what separates 'leverage stacked at risk that cancels' from 'leverage stacked at risk that compounds.' On its own, skew is meaningless without the density component — a perfectly one-sided book on an asset with no at-risk notional isn't a cascade risk, it's just a directional book.

The blend is multiplicative-ish in the cases where it matters. High density × high concentration × high skew is the configuration that earns the highest CSI scores; any one of the three near zero suppresses the index. The exact weights are calibrated against historical liquidation cascades and live in the engine config rather than this page — they're stable parameters, and changing them would invalidate historical CSI series. The component breakdown is exposed on Pro and Elite so a reader can see which of the three is driving the headline number.

What CSI deliberately doesn't do is predict cascades. It measures how primed the topology is for a cascade given the current state — the same way a kindling pile measures fire-readiness without forecasting whether someone will strike a match. The Jacobian endpoint is the right tool for the conditional 'if a shock hit here, how would it propagate.' CSI is the standing-state version of that question.