Allen–Gale Financial Contagion Model

Allen–Gale financial contagion model is the foundational network-economics framework in business and economics theory explaining how cross-holdings of interbank deposits—created to insure regional banks against idiosyncratic liquidity shocks—can act as structural conduits that propagate a small, localized liquidity shortage into a systemic banking crisis across an economy. Introduced by financial economists Franklin Allen and Douglas Gale in their February 2000 Journal of Political Economy article "Financial Contagion", the model extends the single-bank Diamond–Dybvig model of maturity transformation into a multi-region spatial economy connected by an explicit graph of interbank claims. Allen and Gale prove that systemic resilience is a non-monotonic function of interbank network completeness: when the interbank market forms a complete network in which every bank holds diversified, symmetric claims on all other regions, a surprise excess liquidity shock in one region is diluted across all counterparties and absorbed without triggering bank runs outside the origin region; by contrast, when the interbank market forms an incomplete directed cycle (ring network) in which each region holds claims on a single neighboring region, premature fire-sale liquidation of long-term assets in the shocked region forces its creditor bank into insolvency, igniting a domino cascade of financial contagion that bankrupts every bank in the network.

Prior to Allen and Gale (2000), economic explanations of banking panics—from nineteenth-century National Banking Era panics to the 1997 East Asian financial crisis—clustered into two traditions:

1. Macroeconomic Shock Theories: As developed by Gary Gorton (1988) and Allen and Gale (1998), simultaneous runs occur when depositors observe an adverse macroeconomic indicator that impairs the asset values of many banks together.
2. Information-Spillover Contagion: Under information asymmetry regarding loan quality, observing the failure of Bank A leads depositors at Bank B to infer that Bank B holds similar non-performing loans, triggering a run on Bank B even without direct financial claims between the two institutions.

Allen and Gale (2000) isolated a third mechanism: direct balance-sheet contagion through interbank liquidity-sharing networks, even with zero informational spillovers and zero aggregate shocks prior to the crisis. In money markets, banks hold interbank deposits and payment credit lines (analyzed in the Freixas–Parigi–Rochet systemic risk and interbank gridlock model) to reallocate surplus liquidity from low-demand regions to high-demand regions. Allen and Gale showed how the architecture of the interbank network determines whether a small regional liquidity shock remains contained or cascades into economy-wide insolvency.

Timeline, Technologies, and Depositor Preferences

The Allen–Gale economy spans three dates—t = 0, t = 1, and t = 2—and consists of four ex-ante identical regions i ∈ {A, B, C, D}, each populated by a unit continuum of consumers endowed with 1 unit of consumption good at t = 0. Each region has two investment technologies at t = 0:

• Liquid Short-Term Asset (Storage): 1 unit at t = 0 yields 1 unit at t = 1 (or 1 unit at t = 2 if rolled over).
• Illiquid Long-Term Productive Asset: 1 unit at t = 0 yields R > 1 units at t = 2 if held to maturity, but yields a salvage liquidation recovery of r ∈ (0, 1) units if liquidated prematurely at t = 1.

Whereas the Diamond–Dybvig model sets early recovery to r = 1, Allen and Gale assume costly premature liquidation (r < 1), destroying both the future return R > 1 and a fraction (1 − r) > 0 of initial principal. At t = 1, each consumer in region i learns whether they are an early consumer (fraction ω_i, consuming at t = 1) or a late consumer (fraction 1 − ω_i, consuming at t = 2), maximizing expected concave utility:

Negatively Correlated Regional Liquidity Shocks in Normal States

In normal states (S₁ and S₂, each with probability 1/2), liquidity demand fluctuates across regions while aggregate liquidity demand across the four regions remains constant:

• Two regions experience a high liquidity shock ω_H and two experience a low liquidity shock ω_L (0 < ω_L < ω_H < 1), with economy-wide average γ = (ω_H + ω_L)/2.
• In State S₁, ω_A = ω_C = ω_H and ω_B = ω_D = ω_L; in State S₂, ω_A = ω_C = ω_L and ω_B = ω_D = ω_H.

First-Best Decentralization via Interbank Cross-Deposits

Because aggregate liquidity demand equals γ in both S₁ and S₂, a social planner pooling all four regions achieves the first-best Pareto-efficient allocation (c₁*, c₂*) satisfying u′(c₁*) = R · u′(c₂*) with 1 < c₁* < c₂* < R by investing y = γc₁* in the short-term liquid asset and x = 1 − y = (1 − γ)c₂*/(R) in the long-term illiquid asset. Decentralized regional banks implement this allocation by exchanging interbank deposits at t = 0 equal to their Period-1 high-state deficit:

At t = 1, a high-shock region ω_H liquidates its z interbank deposits held in a low-shock region ω_L (which has excess cash (γ − ω_L)c₁* = zc₁*). At t = 2, the low-shock region cashes in its z deposits inside the high-shock region, paying all late consumers c₂*.

While multiple interbank network graphs implement the same first-best risk sharing in the normal states S₁ and S₂, they exhibit contrasting stability properties when hit by a zero-probability (unanticipated) perturbation State S̄.

Three Canonical Interbank Network Architectures

1. The Incomplete Directed Ring Network: Each region holds all z = ω_H − γ units of its interbank deposits in a single adjacent counterparty along a directed cycle: Bank A holds z deposits in Bank B, Bank B holds z deposits in Bank C, Bank C holds z deposits in Bank D, and Bank D holds z deposits in Bank A (A → B → C → D → A).
2. The Complete Symmetric Network: Each region diversifies its interbank claims across all other regions (holding z/2 deposits in each of the two regions with negatively correlated shocks, or z/3 across all three other regions in the symmetric 4-region specification).
3. The Disconnected (Partitioned) Network: Banks exist in isolated pairs (A ↔ B and C ↔ D) with zero interbank claims bridging the two sub-networks.

The Perturbation State S̄ and the Liquidation Buffer b

Suppose at t = 1, instead of S₁ or S₂, the economy realizes the unanticipated perturbation State S̄: regions i ∈ {B, C, D} have the average early withdrawal fraction ω_i = γ, while Region A experiences a small excess early withdrawal shock ω_A = γ + ε (with 0 < ε < 1 − γ).

Regions B, C, and D have enough cash y = γc₁* to pay their γ early consumers at t = 1 if their interbank flows net to zero, whereas Region A faces a net cash shortfall of εc₁* at t = 1. Bank A can raise cash at t = 1 by liquidating long-term assets x at the salvage rate r < 1, provided its Period-2 payout does not fall below c₁* (which would trigger a run by its patient depositors and interbank creditors at t = 1). The maximum liquidation buffer b that a bank with early demand γ can raise at t = 1 without triggering a run by its late depositors is:

Because c₂* > c₁*, the buffer b > 0 is positive whenever r > 0 and vanishes (b → 0) as r → 0. If Bank A's shortfall εc₁* exceeds its buffer, Bank A experiences a bank run at t = 1, liquidates 100 percent of its long-term portfolio x at salvage value rx, and pays every claimant a pro-rata bankruptcy liquidation value q^A < c₁*:

How Contagion Propagates Around the Incomplete Ring (A → B → C → D → A)

When Bank A fails at t = 1 and pays q^A < c₁*, consider the impact on Bank D, which holds all z units of its interbank assets in Bank A:

1. Step 1 (Bank D's Interbank Loss): At t = 1, Bank D owes zc₁* to Bank C and γc₁* to its early depositors, while its z interbank deposits in Bank A pay zq^A instead of zc₁*, creating a cash shortfall Δ_D = z(c₁* − q^A) > 0.
2. Step 2 (Failure of Bank D's Buffer): If z(c₁* − q^A) > b, Bank D cannot cover the shortfall without driving its Period-2 payout below c₁*, so Bank D's late depositors run at t = 1, forcing Bank D into bankruptcy and premature liquidation of x at salvage value rx.
3. Step 3 (Amplification from D to C to B): Because Bank D's interbank asset in Bank A is impaired (q^A < c₁*), Bank D's liquidation payout q^D is strictly lower than Bank A's (q^D < q^A < c₁*):

Consequently, the loss inflicted by Bank D on Bank C, z(c₁* − q^D), exceeds the loss Bank A inflicted on Bank D (z(c₁* − q^D) > z(c₁* − q^A) > b), toppling Bank C (q^C < q^D) and then Bank B (q^B < q^C). Every bank in the ring is driven into bankruptcy and forced to liquidate 100 percent of its long-term investments prematurely.

Why the Complete Network Stops Contagion

Compare the same bankruptcy of Bank A under the complete interbank network, where Bank A's z interbank liabilities are held across all (N − 1) = 3 other regions. When Bank A fails at t = 1, the interbank default loss z(c₁* − q^A) is split across all (N − 1) creditor banks:

As long as this diluted loss falls below each individual bank's liquidation buffer b (z/(N − 1) ≤ b < z(c₁* − q^A)), each creditor bank (B, C, and D) absorbs its share of Bank A's default by liquidating a small fraction of its long-term asset portfolio within its buffer b, keeping its Period-2 payout above c₁*. No late depositors run in Regions B, C, or D, and contagion is halted at the boundary of Region A.

Endogenous Asset-Price Contagion (Allen and Gale, 1998; 2004)

In companion papers (1998, 2004) and Section V of their 2000 article, Allen and Gale replace the fixed salvage technology r < 1 with an endogenous secondary market for long-term assets ("cash-in-the-market pricing"). When Bank A liquidates long-term assets at t = 1 into a market with a limited pool of liquid cash held by arbitrageurs or surviving banks, the market-clearing price P₁ of the long-term asset is determined by the ratio of available cash in the market to the volume of assets sold. A bank failure in Region A depresses the secondary-market price P₁ of long-term assets, reducing the mark-to-market liquidation buffer b(P₁) of all other banks holding the same asset class—combining direct network contagion with indirect price-mediated contagion (related to the Brunnermeier–Pedersen liquidity spiral model and Kiyotaki–Moore credit cycles).

Subsequent Advances: Gai–Kapadia (2010) and Acemoglu–Ozdaglar–Tahbaz-Salehi (2015)

The Allen–Gale (2000) model inspired two major network extensions:

1. Gai and Kapadia (2010, Proceedings of the Royal Society A): Using random-graph percolation theory, Prasanna Gai and Sujit Kapadia showed that financial networks exhibit a "robust-yet-fragile" property: higher connectivity reduces the probability of contagion from small shocks (via the Allen–Gale dilution mechanism), but conditional on a shock large enough to overwhelm the diluted buffers, high connectivity widens the reach of contagion across the graph.
2. Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015, American Economic Review): Daron Acemoglu, Asuman Ozdaglar, and Alireza Tahbaz-Salehi proved a phase transition in financial network stability: when negative shocks are smaller than a threshold ε*, the complete network is the most stable architecture (as in Allen and Gale, 2000); however, when negative shocks are larger than ε*, the complete network drags every bank into default simultaneously, whereas a weakly connected or partitioned network acts as a "financial firewall" that isolates the shocked component while saving the rest of the financial system.

Why do banks form interbank links in the Allen–Gale model if those links can cause contagion?

Regional banks face negatively correlated liquidity shocks in normal times (State S₁ and State S₂): some regions have high early withdrawal demand ω_H while others have low withdrawal demand ω_L. Holding interbank cross-deposits z = (ω_H − ω_L)/2 allows banks to pool regional liquidity and achieve the first-best Pareto-efficient consumption allocation (c₁*, c₂*) without prematurely liquidating long-term assets. Contagion occurs only when an unanticipated aggregate liquidity shock State S̄ hits the network.

Why is an incomplete ring network more fragile than a complete network in Allen and Gale (2000)?

In an incomplete directed ring (A → B → C → D → A), each bank holds 100 percent of its interbank assets z in a single neighbor. When Bank A fails and pays q^A < c₁*, its sole creditor (Bank D) bears the full default loss z(c₁* − q^A), which exceeds Bank D's liquidation buffer b and forces Bank D into bankruptcy, cascading to C and B. In a complete network, Bank A's liabilities are split across all (N − 1) other banks, so each creditor suffers 1/(N − 1) of the loss—small enough to be absorbed within each bank's liquidation buffer b.

What is the role of the premature liquidation recovery rate r < 1 in driving contagion?

When long-term assets suffer steep premature liquidation costs (r → 0), each bank's liquidation buffer b = r(1 − γ)(c₂* − c₁*)/(R) shrinks toward zero, and forced liquidation by a failing bank destroys asset value (paying R → r), amplifying the bankruptcy loss q^D < q^A at every step around the ring.

When can a complete network be less stable than a disconnected network?

As shown by Allen and Gale (2000) for disconnected networks and generalized by Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015), if the initial shock ε in Region A is large enough to exceed the sum of liquidation buffers across all (N − 1) banks combined, even the complete network cannot dilute the loss below b, so all N banks collapse together, whereas a disconnected network lets Region A fail in isolation while preserving Regions B, C, and D.

1. Allen, F., & Gale, D. (2000). "Financial Contagion." Journal of Political Economy, 108(1), 1–33. https://doi.org/10.1086/262109
2. Allen, F., & Gale, D. (1998). "Optimal Financial Crises." The Journal of Finance, 53(4), 1245–1284. https://doi.org/10.1111/0022-1082.00052
3. Gai, P., & Kapadia, S. (2010). "Contagion in Financial Networks." Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 466(2120), 2401–2423. https://doi.org/10.1098/rspa.2009.0410
4. Acemoglu, D., Ozdaglar, A., & Tahbaz-Salehi, A. (2015). "Systemic Risk and Stability in Financial Networks." American Economic Review, 105(2), 564–608. https://doi.org/10.1257/aer.20130456
5. Elliott, M., Golub, B., & Jackson, M. O. (2014). "Financial Networks and Contagion." American Economic Review, 104(10), 3115–3153. https://doi.org/10.1257/aer.104.10.3115