Bank Runs and Gossip Protocols: Why Cascading Failures Look the Same Everywhere

Bank runs and information cascades in distributed systems follow identical mathematical dynamics governed by the SIR epidemic equation: dI/dt = β · I · (N - I) / N. Information spreading and financial contagion produce the same S-curve, critical thresholds, and absorbing states — one converging on coordinated truth, the other on collective collapse. The key difference is not the mathematics but the absorbing state: what the system stabilizes toward when cascading stops.

The SIR epidemic model describes how something spreads through a population of N agents:

dI/dt = β · I · (N - I) / N

Where I is the number of "infected" agents, β is the transmission rate (contact rate times infection probability), and (N - I) is the susceptible population. The shape is always the same: slow start, exponential middle, saturation at the end. An S-curve.

In gossip protocols (SWIM, epidemic broadcast, anti-entropy), I counts processes that have received an update. A node picks a random peer, sends the update, the peer becomes informed. β is determined by the gossip fanout and communication frequency. The absorbing state — where the system stabilizes — is full convergence. Everyone knows the truth.

In liquidity spirals (Brunnermeier-Pedersen, 2009), I counts institutions under distress. A forced seller dumps assets, depressing prices, which erodes collateral at other institutions, forcing more sales. β is determined by leverage ratios and market depth. The absorbing state is market breakdown.

Same equation. Same S-curve. Opposite outcomes.

The structural parallel is striking. Both are contact processes on a network. Both exhibit the characteristic exponential acceleration near I = N/2, where the product I(N-I) is maximized. In gossip, this is the moment when information spreads fastest — half the network is informed and eagerly sharing with the uninformed half. In a liquidity crisis, this is the moment of maximum panic — enough institutions are distressed to drag down the rest, but enough remain standing to be dragged.

The equation is identical but the sign of the outcome depends on what's being transmitted.

Gossip transmits information. Information is non-rivalrous — my knowing something doesn't prevent you from knowing it. When a gossip protocol converges, every node has a consistent view of the world. The system is healthier after the spread completes. You want full infection.

A liquidity spiral transmits distress. Distress is anti-information — it destroys value as it spreads. When Institution A's forced selling drives down the price of assets that Institution B holds as collateral, B hasn't received useful information. B has received damage. The system is worse after every transmission event.

This matters for design. If you're building a gossip protocol, you optimize for fast convergence: high fanout, frequent rounds, redundant transmissions. The faster everyone knows, the better.

If you're regulating a financial system, you optimize for the exact opposite: slow or zero convergence. You want to prevent the cascade from completing. Every policy intervention is an attempt to reduce β or increase the recovery rate.

This framing makes financial regulation legible to anyone who understands epidemiology.

The basic reproduction number R0 in an epidemic is the average number of secondary infections from one infected individual. If R0 > 1, the epidemic grows. If R0 < 1, it dies out. Every public health intervention aims to push R0 below 1.

Financial regulation does the same thing, just with different vocabulary:

The 2008 crisis, in this framing, was an R0 >> 1 event. Leverage was extreme (β was high), counterparty networks were dense (contact rate was high), and there was no quarantine mechanism until the Fed improvised one. The TARP and the emergency lending facilities were, mathematically, a mass vaccination campaign executed mid-pandemic.

Dodd-Frank and Basel III are the post-pandemic public health reforms: mandatory reserves (vaccination), stress tests (surveillance), and the Volcker Rule (reducing contact between commercial and speculative activities).

Both gossip and liquidity spirals share a dangerous property: they're hardest to detect right when intervention matters most.

In the early phase (I is much smaller than N), the spread looks linear and slow. A few nodes informed, a few institutions stressed — no big deal. The exponential character is invisible because I(N-I)/N is approximately I when I is small. It looks like addition, not multiplication.

Then you hit the inflection point around I = N/2, and the acceleration becomes obvious. But by then, half the system is already affected. In gossip, this is fine — you're past the hard part and convergence is inevitable. In a liquidity spiral, you've lost.

Phase        | I/N    | Spread rate  | Intervention window
-------------|--------|-------------|--------------------
Early        | < 0.1  | Slow        | Easy but seems unnecessary
Acceleration | 0.1-0.5| Exponential | Possible but politically hard
Inflection   | ~0.5   | Maximum     | Too late for soft measures
Saturation   | > 0.5  | Decelerating| Damage done, mop up

This is why "early and aggressive" is the universal prescription for both pandemic response and financial crisis management. The math guarantees that waiting for clear evidence of a problem means the problem is already half-complete.

Bear Stearns went from "some liquidity concerns" to "acquired by JPMorgan at $2/share" in ten days. The exponential was running the whole time; it just didn't look exponential until it was too late to stop.

The SIR analogy is clarifying but not perfect. It breaks down in at least three important ways.

Heterogeneous Nodes

Gossip protocols typically assume symmetric, homogeneous agents. Every node has the same fanout, the same probability of being contacted, the same behavior when it receives a message. The math works because the mean-field approximation holds — you can treat the population as a well-mixed soup.

Financial networks are deeply heterogeneous. A community bank and Lehman Brothers are not interchangeable nodes. A single systemically important institution can seed a cascade that no number of small bank failures could trigger. The SIR model with homogeneous mixing underestimates tail risk because it doesn't capture the power-law degree distribution of real financial networks. You need something closer to the BA (Barabasi-Albert) preferential attachment model, where hub failures are qualitatively different from leaf failures.

Continuous vs. Binary State

In gossip, a node is either informed or not. It's a clean binary. Once you know the update, you're done — there's no "partially informed" state that affects your behavior.

Financial distress is continuous. An institution can be slightly stressed, moderately leveraged, or on the verge of insolvency. This gradient matters because it creates feedback loops that the binary SIR model doesn't capture. A 5% price decline might not trigger any margin calls; a 10% decline triggers some; a 15% decline triggers a cascade. The nonlinearity isn't in the epidemic dynamics — it's in the threshold function that maps asset prices to institutional behavior.

Convergence Is Not Guaranteed

Gossip protocols always converge once started (assuming the network is connected and nodes don't crash faster than information spreads). The absorbing state is unique: everyone informed.

Liquidity spirals may or may not reach the breakdown absorbing state. They can stabilize if prices find a level where margin calls stop, if central banks inject liquidity, or if buyers step in to absorb forced sales. The epidemic can "burn out" partway through if β drops fast enough. This is why some financial crises are contained (LTCM 1998, the 2020 COVID liquidity crunch) and others aren't (2008). The equation is the same, but the parameters are dynamic — and policy can change them mid-crisis in ways that biological epidemics can't.

Recovery and Reinfection

In the simplest gossip model, there's no "un-learning." In finance, recovery is possible — an institution can deleverage, recapitalize, and become healthy again. But it can also be re-infected: the same institution can face another liquidity shock before it has fully recovered. This turns the SIR model into something more like SIS (susceptible-infected-susceptible), where endemic equilibria are possible. Chronic financial instability — decades-long periods of repeated crises — maps to the endemic SIS steady state, not the clean convergence of SIR.

Strip away the domain-specific details and you're left with a template that appears everywhere:

1. Local event triggers a state change in one node
2. Contact with neighboring nodes transmits the state change
3. Amplification: the newly changed nodes trigger further changes
4. Cascade or convergence, depending on whether the transmitted signal is constructive (information) or destructive (distress)

You see this in social media (misinformation cascades), supply chains (the bullwhip effect), power grids (cascading blackouts), and distributed databases (partition propagation). The SIR equation or something very close to it shows up every time because the underlying mechanism — local interaction producing global dynamics — is the same.

The lesson isn't that everything is an epidemic. It's that the mathematics of spreading processes is a fundamental tool for understanding any networked system where nodes influence each other. If you can identify β (transmission rate), the threshold function (what causes a node to change state), and the absorbing states (where does the system end up), you can predict the qualitative behavior of the cascade before it happens.

If you're designing a system where participants can "withdraw" — whether that's depositors pulling money, nodes leaving a cluster, sellers flooding a market, or users abandoning a platform — you have a liquidity problem whether or not money is involved. The math doesn't care what's flowing through the network. It only cares about the rate.

• Path Dependence: Why Historical Accidents Shape Markets
• Poverty Traps Are Bistable Systems: A Mathematical Lens on Development Economics
• How Individual Incentives Create Population Equilibrium