The Branching Ratio and Critical Cascades

A forced liquidation sells some assets, which pushes prices down a little, which can force other leveraged holders to sell too, which pushes prices down further. Does this cascade fizzle out after a few rounds, or does it have the potential to spiral without limit? You could try to trace every individual sale and its knock-on effect, but there's a much simpler question that answers it: on average, how many additional forced sales does one forced sale trigger? That single number — the branching ratio — tells you almost everything about whether cascades in this system die out or explode.

An analogy: a chain letter, or a contagious disease

Think of a chain letter where each recipient forwards it to some random number of new people. If each person forwards it to fewer than one new person on average, the chain sputters out within a few generations almost surely, no matter how it started. If each person forwards it to more than one new person on average, the chain has a real chance of growing forever, spreading to an ever-larger population. If the average is exactly one, the chain is on a knife-edge: it eventually dies out with probability 1, but it can take a very long time and grow very large before doing so. This is precisely how epidemics, and financial contagion cascades, behave — it's the same mathematics, whether the "infection" is a virus or a forced sale.

The idea, one symbol at a time

Model the cascade as a branching process: start with one initial event (generation 0). Each event in generation $n$ independently produces a random number of "child" events in generation $n+1$, drawn from the same offspring distribution every time. Let $m$ be the branching ratio — the expected number of children per event:

$$m = \mathbb{E}[\text{number of children per event}] .$$

In plain English: $m$ is the average number of new triggered events each single event causes. The classical branching-process result is a sharp trichotomy:

• If $m < 1$ (subcritical), the cascade dies out with probability 1, and the expected total number of events across all generations is finite: $\mathbb{E}[\text{total size}] = 1/(1-m)$.
• If $m = 1$ (critical), the cascade still dies out with probability 1 eventually, but the expected total size is infinite — cascades can grow very large before stopping.
• If $m > 1$ (supercritical), there is a strictly positive probability that the cascade never dies out and grows without bound.

The formula $1/(1-m)$ is the key quantitative statement for the subcritical case: it says the expected total cascade size, including the seed event, blows up as $m$ approaches 1 from below — exactly the same $1/(1-\rho)$-shaped blow-up seen in queueing.

Worked example 1: a leverage cascade with m below 1

A stress model estimates that each forced liquidation triggers, on average, $m = 0.4$ additional forced liquidations elsewhere (a typical calm-market estimate). Expected total cascade size from one seed event:

$$\mathbb{E}[\text{total size}] = \frac{1}{1 - 0.4} = \frac{1}{0.6} \approx 1.67 .$$

So one forced sale is expected to produce, in total, about 0.67 additional forced sales before the cascade dies out — small and self-limiting.

Worked example 2: the same system under stress

Now suppose a crowded trade and thin liquidity push the same estimate up to $m = 0.9$ (still subcritical, but close to the edge):

$$\mathbb{E}[\text{total size}] = \frac{1}{1 - 0.9} = 10 .$$

The expected total cascade size jumps from 1.67 to 10 — a nearly six-fold increase — from a branching ratio that only moved from 0.4 to 0.9. If stress pushes $m$ to 1.0 or above, the expected size formula itself breaks down (division by zero, or by a negative number), signaling that the cascade is no longer guaranteed to be small; at $m \ge 1$ you need the probability-of-extinction calculation instead of the expected-size formula, and for $m>1$ that extinction probability is strictly less than 1.

What this means in practice

The branching ratio is a standard lens for thinking about self-exciting phenomena in markets: order-flow clustering, volatility clustering (each large move raising the odds of another), and — most consequentially — deleveraging cascades in stressed markets. Risk managers estimate an effective $m$ from historical cascade data (often via self-exciting point process models, see Hawkes processes) and watch how close it sits to 1, since that proximity, not the average level of activity, is what predicts fragility.