Entropy in Financial Contagion Research Michael Stutzer Prof. of Finance University of Colorado, Boulder michael.stutzer@colorado.edu KEYWORDS: contagion, stability, entropy, bankruptcy resolution ABSTRACT: A large literature has evolved to study the reasons and ramifications of default in liabilities networks, e.g. the network of interbank loans, or their customers’ funds’ transfers among each other. Much attention has focused on the prospect for default cascades in which one party’s inability to meet its obligations can prevent its creditors from doing the same, which in turn cause other defaults further downstream. This is a type of financial contagion. The matrix of liabilities between the agents must be estimated. Entropic estimators have been used to estimate the unobserved cells; we argue that this is a reasonable procedure under the circumstances envisioned. In addition, entropy is used to construct new indices of default and contagion severity, and a new index of the inequity inherent in a default resolution. Using the latter in conjunction with the game-theoretic theory of conflict resolution, we demonstrate advantages associated with deviation from the common default resolution procedure that requires a defaulting party to pay its creditors the same proportion of what each is owed. Key Points: • Entropy is used to devise new indices of default severity, contagion severity, and inequity in default resolution procedures • A simple example is used to illustrate these indices, and is bootstrapped to illustrate reasons why entropic liability matrix estimators do not generally bias the default, contagion, and inequity measures in one direction or the other. • Relaxing the usual restriction that the default initiator must pay all its claimants the same proportion (say, 80% of what each is owed) is justifiable within the game theory of bargaining (e.g. the Nash Bargaining solution), and can both increase efficiency while lowering inequity in the default resolution procedure. 1 I. Introduction The standard representation of a payments network starts with a snapshot of gross liabilities owed by each agent (bank, firm, trader, etc.) to each other agent, in the form of a matrix L: L11 L 21 LN 1 L12 L22 LN 2 … L1N … L2 N … LNN in which Lij is an amount that agent i owes to agent j. These are gross rather than net liabilities, so that L ji need not be − Lij ; in fact, all elements are nonnegative. The entries could represent outstanding loan balances, or loan payments that are due, or checks drawn on one bank that must be deposited in accounts at another bank, or payments owed as a result of mutual trading activities, etc. Define the row sum li , column sum a j , and their corresponding shares of the grand totals Li = li / ∑ lk and Aj = a j / ∑ ak . k k We will use the following numerical example for illustration throughout: l #2 #3 # 4 Agent #1 #1 0 0 10 0 10 #2 30 0 20 20 70 Lij ≡ #3 10 30 0 10 50 #4 10 0 20 0 30 50 30 50 30 160 a A .313 .187 .313 .187 L .063 .437 .313 .187 (1) Now suppose for the moment that are no estimation errors. We see that Agent #2 owes l2 = 70 but is owed only a2 = 30 . Without some additional funds to use, it cannot pay all its liabilities, and hence will have to default on some of them. Agent #2 owes 20 to Agent #3, who has no surplus available from its a3 = 50 in assets to pay its l3 = 50 in total liabilities, and hence will also have to default on some payments if it doesn’t receive payment from the defaulting Agent #2. In this way, default by one agent may trigger defaults by others. A cascade of defaults that is triggered by a single default is a type of financial contagion. Here the contagion was triggered by some situation that resulted in Agent #2 owing more in the aggregate than it was due to receive, without outside resources to draw upon. With other matrices, there may be more than one agent initially in default, and those may trigger subsequent defaults. 2 This aspect of credit/payment systems is not only relevant, it may have motivated the advent of bankruptcy law centuries ago. As noted in Kadens (2010, pp.1237-8): “The merchant or trader who relied on credit lived constantly on the edge. The still relatively primitive state of communication, travel, and production meant that he could be sure when he would receive the next shipment or the next payment on which his ability to pay his own creditors depended. His goal was to `synchronize the payments being made to him as a creditor with those he had to make as a debtor’, and this he could never do with complete assurance. As all merchants and traders who depended on credit existed in this state of financial instability, the insolvency of one person who owed significant debts could lead to the failure of many others.” One might conjecture that financial contagion is less likely when the liabilities are more evenly distributed. The intuition is that a default by one agent will cause many agents to lose small amounts, insufficient to cause any of them to default on their own obligations. In order to study this conjecture, one needs to define what is meant by “liabilities are more evenly distributed”. The entropy defined in section II can be used for this purpose. Then one needs to define a procedure for resolving defaults. Is an agent who is stiffed by another permitted to renege on its own obligations to it? Can a bankruptcy judge determine who it will renege on? Following Elimam, et. al. (1996) and Eisenberg and Noe (2001), the literature has focused on a default resolution process in which defaulting agents are required to proportionally renege on all their respective creditors. In their words (op.cit., p.239) “all claimant nodes are paid by the defaulting node in proportion to the size of their nominal claim on firm assets.” In other words, after any default cascade has ended, an agent that can pay only θ % of its total liabilities must pay exactly θ % of the funds it owes to each of its creditors. 1 The equilibrium resolution procedure is described in section III, and used to formulate equilibrium, entropic indices of both default and subsequent contagion severity. In section IV, we show why there needn’t be any particular relationship between the entropy of a liability matrix and these default and contagion severity indices is explored in section IV. Perhaps fairness considerations might favor different default resolution procedures in some circumstances, and laws could be changed to permit or even encourage bankruptcy courts to implement different procedures on grounds of fairness. In section V, an entropy-based index of inequity is proposed. Section VI demonstrates beneficial effects on both default severity and inequity that stem from relaxing the proportional payment rule for an agent who initiates a default cascade. Section VII both reviews and adds new results to the significant game-theoretic results that support alternative bankruptcy resolution procedures, because the theorems previously developed are for single debtor vs. multiple claimant situations, and are incorrect in financial networks. Section VIII concludes. 1 The Proportional Rule was retained when Rogers and Veraart (2013, p. 884) extended the Eisenberg and Noe (op.cit.) framework “by introducing costs of default if loans have to be called in by a failing bank” (p.882). 3 II. The Entropy of the Liabilities Matrix The survey paper by Upper (2011, sec. 4.2) describes a widely-used procedure to define an entropy of the liabilities matrix L is to normalize it by dividing each of its cells by the grand total of all cells, i.e. define Pij = Lij / ∑∑ Lij , and compute the Shannon entropy of the normalized i j matrix H = −∑∑ Pij log Pij . By adopting the convention 0 log 0 = 0 , so cells with zeros, e.g. i j those along the diagonal, contribute nothing to entropy, and hence are omitted from the double sum, i.e. we calculate H = − ∑ Pij log Pij In our example (1), the Shannon entropy H = 2.10 . ij ;i ≠ j Suppose that all cells off the diagonal are unknown, but that the column and row sums of L are observed. This situation was faced by early researchers with access to financial reports that listed total liabilities and assets of agents without breaking out the bilateral specifics. The unknown cells have been estimated by maximizing this entropy subject to the constraints that row and column sums have fixed values (typically known by observation). That is: maxPi≠ j − ∑ P log P ij ;i ≠ j ij ij subj. to= Aj= : ∑ Pij L= ; ∑ Pij 1 i ; ∑ Pij j ≠i i≠ j (2) ij ;i ≠ j The problem re-defined with sums over all cells has the solution Pij = Li Aj , i.e. the joint distribution with maximum entropy would be the product of the marginals,, as-if we had assumed the distribution of liabilities was independent of the distribution of assets. 2 But that Pii Li Ai ≠ 0 . In accord with this problem, Upper and Worms would imply the counterfactual= (2004) reformulate the problem as one of finding the joint distribution as close to the product of the marginals as possible (measured by the entropy of the former relative to the latter) when Pii = 0 . Formally, one minimizes the mutual information subject to the known row and column totals and the constraints Pii = 0 : min Pi≠ j Pij ∑ P log L A ij ;i ≠ j ij i j .t. ∑ Pij L= ; ∑ Pij 1 s= Aj= i ; ∑ Pij j ≠i i≠ j (3) ij ;i ≠ j The objective function in (3) (i.e. mutual information) is seen to be the relative entropy of the offdiagonal probabilities Pij with respect to the product measure Lij . Using illustrative example (1), suppose that we did not observe values of the cells but did observe the row totals li and 2 This is a consequence of Theorem 2.6.6 in Cover and Thomas (1991, p.28). 4 column totals a j needed to calculate the marginal probabilities Li and Aj listed in (1). The Excel Solver® quickly found a solution to (3), from which the liabilities Lij = Pij / ∑ li are i recovered. Rounded to two decimal places (causing some minor adding-up errors), they are 3: #1 #2 #3 #4 l Agent #1 0 3.13 4.69 2.17 10 #2 22.36 0 32.56 15.08 70 18.89 18.36 0 12.74 50 #3 #4 8.75 8.50 12.74 0 30 50 30 50 30 160 a A .313 .187 .313 .187 L .063 .437 .313 .187 (4) Comparing (4) to (1) illustrates how the entropic estimator (3) spreads the liabilities more evenly. Three off-diagonal cells in (1) were zeroes. None of them are zero in (4). In (1), Agent #1 owed all liabilities to a single agent (#3). In (4), that was spread across all three other agents. The mutual information in (4), i.e. the value of the objective function in (3), is only 0.288. The (higher) mutual information in (1) is 0.466. This estimator is rationalized by appealing to the principle that one should incorporate all information that is known, and nothing else. Here, what is known are the respective row and column sums, and that diagonal elements are zero (no agent owes payments to itself). Had the diagonal elements also been unrestricted, the solution to (3) would indeed be Pij = Li Aj achieving the global minimum mutual information equal to zero, and hence Lij = Pij / ∑ li . If more i information is known, e.g. some of the individual cells’ values are observed, we need only subtract them from their respective row and column totals, and then drop the corresponding probabilities from being estimated by (3). Upper (op.cit.) notes that such entropic estimators will not place zeros in any off-diagonal cell, counterfactual with evidence garnered from more complete datasets. Interbank lending has a tiered structure, in which numerous, smaller banks are linked to fewer, much larger “core” banks, which in turn are linked to each other. In this “two-tiered” structure, the large number of small banks are not linked closely with each other, introducing a degree of sparseness in L which would not be reproduced by the above procedure. But he also notes that the analyst could introduce additional constraints on the problem to better represent these aspects. One could make use of statistics that characterized more complete datasets in analogous circumstances. Again, the foundation of the resulting estimators is to incorporate known information – but nothing else – into the constrained minimization. 3 (4) would still result had we maximized the constrained Shannon entropy rather than minimized the constrained mutual information. 5 III. Entropic Indices of Default and Contagion Severity In calculations concerning default, we first must consider the simultaneity problem in the liabilities network : an agent owes funds to others, but in turn is owed funds by them. Suppose we assume the proportional payment rule requiring each bankrupt agent i to pay a maximal proportion 0 ≤ θi < 1 of its separate liabilities to each of its creditors, i.e. a pro rata distribution, when it can’t fully pay all of them (if it can fully pay of them, then it does not default, i.e. θi = 1 ). How do we know that an equilibrium clearing vector = θ (θ1 , θ 2 , …, θ N ) of recovery shares exists? In other words, how do we know that the proportionality rule can be applied to all agents simultaneously? Eisenberg and Noe (op.cit.) proved that a clearing vector does exist, provided conditions under which it is unique, and showed that a linear programming problem can be solved to find a maximal clearing vector. We follow the lucid exposition of Demange (2015) to define the clearing vector problem and its solution. The solution specifies that agent i pays agent j X ij = θi Lij , so that the aggregate of payments from agent i will be ∑X j ≠i ∑X j ≠i ji ij = θi li , while the aggregate of payments to agent i will be = ∑ θ j L ji . To focus sole attention on the role of the liabilities matrix, in what follows I j ≠i assume that the agent has no exogenous resource ei available to pay shortfalls in the event that the aggregate of payments made to agent i are insufficient to cover its aggregate liabilities. 4 So the equilibrium requirement is that a vector θ between 0 and 1 satisfies the linear inequalities: θi li − ∑ θ j L ji ≤ 0; i = 1, …, N j ≠i Eisenberg and Noe (op.cit.) prove that a maximal equilibrium clearing vector θ* must maximize any increasing function of θ , in particular it could maximize the aggregate liabilities paid throughout the network: θ* = arg maxθ1 ,…,θ N ∑θ l i i i s.t. θi li − ∑ θ j L ji ≤ 0; i = 1, …, N (5) j ≠i 4 Letting agents have exogenous funds to cover defaults only complicates the issues addressed here. In actuality, rules or laws must be mutually or externally enacted and enforced to ensure that agents maintain fixed levels of collateral that can be assigned to cover defaults, so such analyses will be situation-dependent. Also, if such collateral requirements are high enough, there will be no initial bankruptcies, much less contagion. When collateral requirements are less than that, simulations of default and contagion would be dependent on both the magnitudes and the distribution of the collateral, complicating our goal of understanding the relationships between the structure of the liability matrix L, and the default, contagion, and inequity measures developed herein, as well as the alternative default resolution procedures analyzed herein. That understanding is enhanced by assuming situations in which default is not a rare event, as it will be when assumed collateral is high enough. Readers interested in estimates for a particular financial network can easily modify the analysis herein to incorporate that network’s distribution of assignable collateral. 6 In our illustrative example (1), the Excel Solver® quickly found the solution θ* = (1.0, 0.148, 0.344, 0.213) . We see that Agent #1 does not default, and hence pays the full 10 that it owed. As described earlier, Agent #2 had to default, because it owed 70 but was scheduled to receive only 30 from others. In equilibrium Agent #2 only pays θ 2 = 14.8% of its l2 = 70 total liability. This triggered a cascade in which both Agent #3 and Agent #4 defaulted, respectively paying only 34.4% and 21.3% of their total liabilities. They both had originally owed amounts equal to what was due from others, and hence were vulnerable to the impossible situation of Agent #2. Only Agent #1 avoided default, due to its healthy originally situation in which it owed 40 less than was due from others. After clearing, the realized payments X ij are: Agent #1 #2 X ij ≡ #3 #4 * a A* #4 l* L* 0 10 .228 2.951 10.328 .235 3.443 17.213 .392 2.131 0 4.262 0 6.393 .146 10 10.328 17.213 6.393 43.934 .228 .235 .392 .146 #1 #2 0 0 4.426 0 3.443 10.328 #3 10 2.951 0 (6) Examination of the Proportional Rule’s resolution payments matrix (6) reveals a number of the rule’s properties. We see that the defaulting Agents #2, #3, and #4 all pay out the amounts they each receive, i.e. l *i = a*i for all of them, so the corresponding constraints in (5) are binding. Thus the solution to (3) incorporates the common bankruptcy provision that receivables of defaulting agents are fully paid out to creditors; nothing is left for others. While (1) shows that 40 offset by the bankruptcy triggering Agent #2’s Agent #1 had positive “net worth” a1 − l1 = −40 , the default clearing process wipes those out, i.e. negative net worth a2 − l2 = l *1 − a*= 10 − 10 and l *2= − a*2 10.328 − 10.328 so the first and second constraints in (5) are 1 binding. While ex-ante aggregate liabilities (and hence aggregate assets) owed both totaled 160, ex-post clearing the (objective function) total liabilities paid is only 43.934. Moreover, the distribution of liabilities and assets (i.e. the vectors L and A) have changed. Note that the fraction of liabilities paid ex-post by defaulting Agents #2 and #4 fell from the fractions they owed ex-ante, while the opposite occurred for Agents #1 and #4 – despite the default of Agent #4. How do we quantitatively measure default and contagion? First, we define measures of default and contagion frequency. The total fraction of defaulting agents, both those initially defaulting and those whose subsequent defaults are triggered by them, is denoted D = card {θi* ;θi* < 1} / N 7 (7) In our example, that is D = 3/4. While this captures the frequency of default, it does not measure the frequency of contagion. Two of the four agents were in the default cascade, and we define Dc =2/4 as the contagion analog of (7). The severity of default and contagion is an additional measure that augments the mere number of agents defaulting, because the agents who default may be relatively large and consequential. I propose a new, entropic measure of default severity. Recall that the shares of total liability originally owed by the agents is the distribution L = (l1 / ∑ l j , …, lN / ∑ l j ) . After clearing, the j j vector of fractions actually paid is denoted Q ≡ (θ L , …, θ LN ) . Note that the * 1 1 * N elements of Q are also nonnegative, and are bounded above by the corresponding elements of L, but do not sum to one unless there is no default (i.e. θ* ≡ 1 ). We can still define the nonnegative, relative entropy-like Default Severity Index: L S ≡ ∑ Li log i i Qi 1 = ∑ Li log * ≥ 0 i θi (8) which is zero if and only if all θi* = 1 , i.e. no agents default. We see that an agent contributes more to the default severity index (8) when the agent’s share of total liabilities owed is higher and the fraction of its liabilities paid is lower, in accord with the connotation of severity. In addition, we have: Proposition 1: Minimizing the Default Severity Index (8) can serve as a substitute objective function in (5) to produce the maximal clearing vector θ * . Proof: Eisenberg and Noe (op.cit.) proved that constrained maximization of any increasing function of the vector θ produces the maximal clearing vector θ * . Hence it can also be produced by constrained minimization of any decreasing function of θ , like (8). QED The aforementioned properties make the Default Severity Index (8) is a useful measure of default consequences. In our example (1), S = 1.46. We can also separately measure the severity of any contagion that is triggered by a default. We just calculate the fraction of (8) attributable to subsequent defaults triggered by initial default(s). In the case of (1), the defaults of Agents #3 and #4 were triggered by the default of Agent #2. So we calculate the Contagion Severity Index: Sc ≡ L3 log(1/ θ3* ) + L4 log(1/ θ 4* ) ∑ Li log(1/ θi* ) i which is 42.7% in Example 1. IV. Is the Mutual Information Related to Default and Contagion? 8 (9) What are the relationships, if any, between the Shannon entropy or mutual information in the liabilities matrix and these measures of default severity? The issue originally arose when estimating the full liabilities matrix knowing only its row and column totals. Note that matrices (1) and (4) have the same row and column totals. The fear was that entropic estimation of the unknown liabilities matrix might introduce a predictably signed estimation bias when estimating default and contagion severity from the estimated liabilities matrix. If lower mutual information is achieved by spreading the liabilities out across agents, default by an agent i may adversely affect more agents, but perhaps each of those agents can absorb relatively small losses better than in matrices with higher mutual information, in which the defaulting agent’s liabilities could (but don’t have to) be more concentrated. This would suggest a possible positive bias indicated by a positive relationship between mutual information and default or contagion severity, at least when comparing liabilities matrices with the same row and column totals. But suppose the more evenly-spread liabilities in the matrix with lower mutual information are larger than what the other agents can absorb. Then contagion may be worse than had the large liabilities been concentrated on just one or a few other agents. This suggests a possible negative relationship between mutual information and default or contagion severity, among matrices with the same row and column totals. In his survey, Upper (op.cit., sec. 4.2) summarizes the findings of earlier studies. Such studies used different simulation methodologies and different measures of default and contagion severity than will be used herein. He reported mixed findings among those studies. A complex set of simulation experiments (conducted post-Upper) was conducted by Sachs (2014). She randomly generated liability matrices in which all banks had equal total assets. She found that the relationship between entropy and her measure of contagion (not based on the equilibrium clearing vector) depended on the sparsity of the liability matrix. What is the connection in our example, using our equilibrium clearing vector-based definitions? The Shannon entropy of (1) is 2.10, while that of (4) is 2.8, while the mutual information in (1) is 0.466 vs. 0.288 for the minimal mutual information matrix (4). We saw that the lower mutual information in (4) was indeed achieved by spreading liabilities more evenly across cells. While in both cases one agent initially defaults and triggers subsequent defaults of two other agents, the Default Severity Index (8) rose to S = 1.56 from 1.46 when the liabilities matrix in (4) was substituted for the matrix in (1). We see that in this example, the mutual information (Shannon entropy) was negatively (positively) related to the Default Severity Index (8). The Contagion Severity Index (9) behaves similarly. It increased a bit, from 43% using matrix (1) to 45% using the minimum mutual information matrix (4). Hence in this example, the mutual information in the liabilities matrix and measures of default and contagion severity are negatively related, despite the somewhat more evenly distributed liabilities occurring in the lower mutual information matrix (4) with the same row and column totals. Hence, we see that intuition about the effects of spreading-out liabilities can be misleading. 9 To generate more evidence, a simple, easily replicable way to simulate liabilities matrices is now adopted. 5 First, we permute the off-diagonal elements in (1), to produce other possible liabilities matrices with identical numbers in them. Note that permuting the off-diagonal elements will result in matrices with the same Shannon entropy, because permutation of matrix elements will permute the labels of the various Pij , but won’t change the sum of products defining the entropy. Hence there cannot be any relationship between the Shannon entropy and default or contagion measures produced from matrices resulting from these permutations. But because the row and column totals will not be preserved by these permutations, the mutual information of these matrices will differ, and hence in principle can be related to default and contagion severity. In order to provide evidence based on comparisons to matrices with identical row and column totals, each matrix produced by permutation is paired with the minimum mutual information matrix produced from its row and column totals, i.e. its version of (4). Another advantage of this procedure is that it fixes the network’s total liabilities (and hence network total assets) in each pair to be the same as in the base example. Example 1 has total liabilities of 160. 500 paired matrices were produced by permuting the offdiagonal cells to produce an analog of (1), and then using each to produce the corresponding analog of (4). In none of the 500 pairs does the minimal mutual information matrix result in more contagion frequency than its higher mutual information counterpart. 77% of the pairs have the same contagion frequency, 19% have one fewer contagion default, while the rest have two fewer contagion defaults. But the total number (i.e. initial + contagion) defaults are the same in each pair, i.e. less contagion defaults are countered by more initial defaults. Hence, among matrices with the same row and column totals, it would be misleading to infer that total default frequency is lower when the mutual information is lower. Moreover, there is no relationship between the change in mutual information and the change in default severity (8) or contagion severity (9). To illustrate, Figure 1 plots the (positive, by construction) decrease in mutual information between each pair and the corresponding change in Default Severity Index (8). Across the 500 pairs, the mean change in default severity was near zero (0.03) with a standard deviation over 5 times that. The scatterplot shows little connection between the two, with a correlation coefficient of -11%. Because there is no theoretical reason to believe the connection is linear, perhaps Kendall’s rank correlation τ is a better measure of correlation, but it is only -3.6%. Similar findings occur when substituting contagion severity (9) for default severity (8). On average over the pairs, it dropped a small amount (0.06), but with a standard deviation more than twice that. The Kendall τ rank correlation between the decrease in mutual information and change in contagion severity is only 9.6%. Now suppose we do not wish to hold anything constant except the actual numbers appearing in Example 1, i.e. instead of permuting the elements, we produce a different 500 pairs by bootstrapping the off-diagonal elements in (1). That is, we sample the off-diagonal elements in (1) with replacement rather than without, again pairing each resulting matrix with its minimum 5 The more complex simulation methodologies used in the cited papers are more extensive, but not as easily replicable by outside parties, and still rely on specific examples or heroic assumptions, e.g. Sachs’ (op.cit.) assumption holding fixed the distribution of assets across agents. 10 mutual information estimate having the same row and column totals. In contrast to the permutations, this will produce liabilities matrices with different total network liabilities. We still found no connection between the change in mutual information and the change in either the severity or contagion indices. The Kendall τ correlations were all below 5%. Corroborating visual evidence in seen in Figure 2, a scatter plot of the positive decrease in mutual information and the corresponding change in default severity for each of the 500 pairs. Why Isn’t Mutual Information More Closely Related to Default and Contagion? The mutual information is a measure of the dependence between the row proportions vector L1 , …, LN and the column proportions vector A1 , …, AN considered as two random variates determined by a random liability matrix L . While the mutual information is zero when the row and column proportions are independent, it is always positive regardless of whether the dependence (e.g. correlation) is positive or negative. But there is a signed dependency measure that is closely connected to the severity and contagion indices. That characteristic is the rank correlation between agent liabilities and agent assets. The (sound) intuition is that default and contagion will be more severe when agents with relatively high total liabilities have relative low total assets from which to pay them. Because there is no reason to expect a linear connection, we surmise that the Kendall rank correlation τ L , A between the agents’ respective shares of liabilities and assets will be negatively related to the Default Severity and Contagion Severity Indices. Evidence for that is now provided. Figure 3 uses the same 500 bootstrap replications of Example 1 to illustrate the negative relationship between the Default Severity Index S and τ L , A . 6 Figure 4 shows that the Contagion Severity Index Sc in (9) is positive much more often when that rank correlation is negative than when it is positive, in accord with the above intuition. The mutual information of the liabilities matrix is a measure of dependence between the row (liability) and column (asset) totals. But isn’t a signed measure of dependence, like the Kendall τ rank correlation. We have Proposition 2: Figures 3 and 4 show that negative dependence (i.e. τ L , A < 0) between agents’ liabilities and assets is directly related to default and contagion. While mutual information is a measure of that dependence, it is nonnegative so can be high when dependence is positive as well as negative. This is why there is no definitive link between mutual information and default and contagion. Because the two vectors have only 4 elements apiece, Kendall’s τ can only assume a small number of values. This accounts for the discreteness of the horizontal axis values plotted in Figures 3 and 4. 6 11 V. Inequity of Bankruptcy Resolutions: An Entropy Measure Given the strict proportional payment clearing rule, the clearing vector θ * minimized the default severity, which could be taken as a measure of efficiency in bankruptcy resolution. But what about fairness? There can be dimensions to fairness of default resolution other than preventing defaulting agents from paying different percentages of what it owes to its creditors. In example (1), the agents’ respective shares (subject to a bit of rounding error) of the 160 in total liabilities originally owed are L = (6.3%, 43.7%,31.3%,18.7%) . Ex-post clearing, (6) shows that their corresponding shares of the 43.93 in total liabilities paid are L* = (22.8%, 23.5%,39.2%,14.5%) . How fair is it that these share vectors are different? Perhaps the fairest outcome would be if the distribution of ex-post shares paid L* = L , which is the distribution of ex-ante shares owed. Barring the feasibility of that, it is proposed that the entropy of L relative to L* be used to measure the inequity of the bankruptcy resolution. That is, our measure of inequity I is: I = ∑ Li log i Li L*i (10) Plug the last columns of (6) and (1) into (10) to calculate I = 0.168. A little algebra shows that the Inequity Index (10) and Default Severity Index (8) are related by I= S + log(∑ θi* Li ) ≡ S + log(∑ θi*li / ∑ li ) . i i (11) i That is, the Inequity Index is the Default Severity Index, plus the log of total liabilities paid as a fraction of what was originally owed. We see that it embodies a tradeoff between the two desiderata of low default severity and high total liabilities paid (the objective function in (5)), when the latter is weighted by its log. 7 As an alternative to maximizing ∑θ l in (5) (or equivalently, minimizing the Default Severity * i i i Index (8)), we can consider finding a clearing vector that minimizes the Inequity Index (10). In Example (1), it turns out that the respective solutions are the same. To see this, we first re-write the strict proportional payment clearing constraints in (5) as the system of linear inequalities M θ ≤ 0 . Eisenberg and Noe (op.cit., Theorem 1) prove that even if there is more than clearing vector θ , the negative slack (if any) in each inequality i (which is li* − ai* ) is the same, and there is no slack for any defaulting agent (i.e. θi < 1 ). For the clearing vector θ * solving (5) and used to produce (6), only Agent #1 does not default, and that constraint also has no slack (like the three defaulting agents’ constraints). Hence even if there were multiple clearing vectors, they must be solutions of the linear equations M θ = 0 , i.e. they must be in the null space of M. In example (1), this null space turns out to be one dimensional, 7 If the latter weren’t weighted by its log, there would be no tradeoff, because Proposition 1 shows that constrained minimization of the default severity index produces the same clearing vector as constrained maximization of total liabilities paid. 12 spanned by the vector (61/ 13,9 / 13, 21/ 13,1) . Thus a clearing vector must be proportional to this vector. Because the constant of proportionality cancels when one computes the vector L* used in (10), any clearing vector yields the same value of (10), i.e. I = 0.168 . VI. Would Deviations from Strict Proportionality Be Desirable? Perhaps deviation from the strict proportional payment rule would permit clearing with less default severity (8) and/or less inequity (10). In Example (1), suppose the cascade-triggering Agent #2 is permitted to pay different fractions of what it owes to the other agents. Each other agent is still required to pay a constant proportion. Denote a resulting (not necessarily proportional) clearing vector by θ = (θ1 , θ 21 , θ 23 , θ 24 , θ3 , θ 4 ) , in which case the clearing vector must satisfy the modified inequalities from (5): 0 −10 −10 10 −30 0 0 30 20 20 −30 0 −10 0 −20 0 50 −20 0 0 −20 −10 30 0 θ1 θ 21 θ 23 ≤ θ 24 θ3 θ 4 0 0 0 0 0 0 (12) This also necessitates modifying the objective function in (5) to be the inner product of the 6component θ with the vector (10,30,20,20,50,30), whose 2nd – 4th components reflect the possibility of separate fractions payable by Agent #2 to Agents #1, #3, and #4. The constrained maximum of that subject to (12) is θ max = (1, 0, 1, 7.1%, 71.4%, 28.6%) . From θ max we see that Agent #1 still pays 100% of the 10 that it originally owed. Agent #2 pays none of the 30 it owed to Agent #1, while paying 100% of the 20 it owed to Agent #3, and 7.1% of the 20 it owed to Agent #4, summing to 21.429 of the 70 it originally owed the others. Agent #3 pays 71.4% of the 50 that it originally owed, and Agent #4 pays 28.6% of the 30 that it originally owed. Aggregate to find that of the 160 originally owed by them, the agents respectively paid Q = (6.3%, 13.4%, 22.3%, 5.4%) of it. They originally owed shares of that 160 equal to L seen in the last column of (1). Using these vectors Q and L, the Default Severity Index (8) is 0.858, much lower (i.e. better) than its value with proportional clearing (1.46). The Contagion Severity Index (9) is 39.6%, which is also lower (i.e. better) than its value with strict proportional clearing (42.7%). Hence, we see that both default and contagion severity improve when Agent #2 is allowed to pay different percentages of the liabilities it owes to the other agents, i.e. when Agent #2 is permitted to violate the Proportional Rule. The liabilities X ij paid after this modified clearing procedure are reproduced below: 13 Agent #1 #2 X ij = #3 #4 a max max A L* 0 0 10 0 10 .132 0 0 20 1.429 21.429 .283 7.143 21.429 0 7.143 35.714 .472 2.857 0 5.714 0 8.571 .113 10 21.429 35.714 8.571 75.714 .132 .283 .472 .113 #1 #2 #3 #4 l max (13) Total liabilities of 75.714 are paid in the modified clearing equilibrium, vs. 43.93 paid under the t Proportional Rule. So that performance measure is better, too. But what happens to inequity, i.e. how close is (10) to zero? Plug the last columns of (13) and (1) into (10) to calculate that I = 0.110 , compared to I = 0.168 under the proportional clearing rule. Hence inequity is lowered (i.e. improved) under the modified clearing procedure. Treatment of the initially defaulting Agent #2 seems reasonable, too. Comparing (13) to (6) shows that the initially defaulting Agent #2 pays far more to Agent #3 and only somewhat less to Agents #1 and #4, resulting in total payments close to 11 higher than occurred under strict proportionality. In summary, permitting the cascade-triggering Agent #2 to deviate from the proportional payment rule resulted in less default and contagion severity and less inequity in recovery. How common is this phenomenon? Among our 500 bootstrapped liability matrices, there are 39 in which the only initial default is by Agent #2. In each case, the default severity index is lower when the proportional payment requirement is relaxed for Agent #2 in the way described above. The decline in default severity averaged 31%. In no case did the default by Agent #2 trigger more defaults. In 21 of the 39 cases the number of triggered defaults remained the same, one less agent defaulted in 14 of the cases, while two less agents defaulted in the other 4 cases. Hence contagion frequency was never greater, and frequently was lower. Not surprisingly, contagion severity also declined in 37 of the 39 cases (the other two cases already had very low contagion severity), by an average of 68.7%. The inequity index (10) also declined in 32 of the 39 cases. The average decline was a modest 19.1%. We have Proposition 3: The decrease in default and contagion severity enabled by relaxing the proportional payment requirement for the initial defaulting Agent #2 is not occurring at the expense of higher inequity. The upshot of this simulation is that the previous calculations with Example 1 are not misleading. Moreover, the null space of the matrix in (12) has dimension 3. As such, it is possible that constrained minimization of the inequity index (10) or (11) will yield a clearing vector different than the clearing vector θ max = (1, 0, 1, 7.1%, 71.4%, 28.6%) by maximizing aggregate liabilities paid when the proportional payment rule was relaxed for Agent #2. The constrained minimization of (10) yields θ min = (1, 0, 68.5%, 25.1%, 62.4%, 37.6%) that achieves the minimum I = .098 versus the slightly more inequitable I = 0.110 achieved by maximizing 14 aggregate liabilities paid. The liabilities paid, after clearing that minimizes the Inequity Index, are reproduced below: Agent #1 #2 #3 #4 l* 0 0 10 0 10 #1 #2 0 0 13.704 5.026 18.730 X ij = #3 6.243 18.730 0 6.243 31.217 #4 3.757 0 7.513 0 11.270 * 10 18.730 31.217 11.270 71.217 a A* .140 .263 .438 .158 L* .140 .263 .438 .158 (14) Comparing (13) to (14), we see that the tradeoff for achieving the modest decrease in inequity (from .110 to .098) is a modest loss of 75.714 - 71.217= 4.5 in aggregate liabilities paid corresponding to a modest increase in default severity from 0.858 to 0.909. Comparing (13) to (14), we see that permitting the cascade-triggering Agent #2 to deviate from the proportional payment rule resulted in different bankruptcy resolutions, depending on whether one wants the highest feasible aggregate payments (equivalently, the lowest feasible default severity index) or the lowest feasible inequity. VII. Is the Proportional Rule Really Equitable? Game-Theoretic Insights Results in the previous section raise questions about the Proportional Rule’s merit. Fortunately, there is a large literature that examines the normative basis for the Proportional Rule and plausible alternative rules. Thomson (2003 and 2015) has authored two surveys of this large literature. The literature considers a single agent i who owes amounts Lij > 0 to N separate claimants, but only has resources ai < ∑ Lij ≡ li available to pay them. Generalization of this j ≠i set-up to our networks -- in which debtors and claimants are not mutually exclusive -- has not heretofore been done, but will be later in this section. Theorem 1 in Thomson (2003) summarizes results in Dagan and Volij (1993) that use gametheoretic bargaining models to characterize both the Proportional Rule and the very plausible alternative of equalizing dollar (rather than percentage) payments across the sole defaulter’s claimants. Thomson refers to the latter as the Constrained Equality (CE) Rule. Using X ij ≤ Lij to denote the payments made by the sole defaulting agent i after resolution, the CE Rule identifies a constant Ci such that : X ij min = ( Lij , Ci ) s.t. ∑ X ij ai and X ij ≤ Lij (CE) (15) j ≠i as opposed to our previously modeled Proportional Rule, which identified a constant θi such that X ij θ= ai and X ij ≤ Lij . In (15), the min operator ensures that no claimant is = i Lij s.t . ∑ X ij j ≠i 15 paid more than that claimant is owed (i.e. X ij ≤ Lij ), and that all claimants j are paid the same amount Ci when their respective inequalities are strict, i.e. claimants not made whole receive the same dollar (rather than percentage) payment. 8 In our example (1), the literature applies when considering the plight of Agent i = 2 in isolation, = L21 30, = L23 20, and = L24 20 but only has a2 = 30 available to pay them. The CE who owes = X 23 = X 24= C= 10 , i.e. the claimants split the Rule (15) resolves the situation by setting X 21 1 defaulter’s assets equally. Had a2 = 65 instead, the CE resolution would have been X 21 = 25, X 23 = X 24 ≡ C1 = 20 , because the third and fourth agents can’t be paid more than they are owed (i.e. 20 apiece). Consider the bargaining game between the sole defaulter, denoted i, and the multiple claimants j ≠ i who cannot all be made whole. A constrained feasible vector of payments X i satisfies X ij ≤ Lij and ∑ X ij = ai 9. In the absence of a bargaining solution, it is assumed that j ≠i each claimant receives nothing. A Nash Bargaining solution is a feasible vector of payments that maximizes the product of claimants’ utility gains ∏ [U j ( X ij ) − U j (0)] , where U j is a (not j ≠i necessarily strictly) concave utility function. A weighted Nash Bargaining solution for w weights wij maximizes ∏ [U j ( X ij ) − U j (0)] ij . Theorem 1 in Thomson (2003) cites the Dagan j ≠i and Volij (1993) results that when utility functions U j are linear with U j (0) = 0 , (i) the CE Rule (15) is a Nash Bargaining solution, while (ii) the Proportional Rule is a weighted Nash Bargaining solution when the weights wij = Lij . The former result reflects the presumption that the creditors have equal bargaining weight over the default resolution process, while the latter reflects the presumption that those who are owed more have logarithmically more weight in proportion to what they are owed. On this basis and others summarized in Thomson (2003), the CE Rule deserves consideration as an alternative to the Proportional Rule heretofore analyzed. As noted, the formulations in Dagan and Volij (op.cit.) use linear utility functions U ij = X ij , while our formulations above assumed concave utilities; moreover, their proofs are informal. It is insightful to see the following formal proof of this more general result. Levinthal’s (1918) history of early bankruptcy law notes that ancient Jewish law required equal dollar (rather than percentage) payments to creditors, subject to the constraints that this would not compensate any creditors more than they were owed (op.cit, p.234). Indeed, Thomson (2003) writes that this was advocated by Moses Maimonides, the influential 12 th Century Jewish Sage. This is not what is sometimes called the “Talmud Rule”, see Thomson (2003, p.256). 8 9 Replacing the equality with the less than or equal to sign is possible, but the desideratum of Pareto Optimal payments implies that rules will result in payments that sum to ai . 16 Proposition 4: When there is a single defaulting agent i , and claimants j ≠ i have concave utilities U j ( X ij ) with identical threat points U j (0) = 0 , the Nash Bargaining solution is a CE rule. Proof: The Lagrangian for the problem is ∏ = j ≠i U j ( X ij ) − ∑ λ j ( X ij − Lij ) − γ ∑ X ij − ai j ≠i j ≠i We assume U j ( X ij > 0) > 0 and that U j ' > 0 and U "j ≤ 0 . Then one can replace the objective function U ′j ( X ij* ) * ij U j (X ) ∏ j ≠i U j ( X ij ) := log ∑ U j ( X ij ) in the Lagrangian. The first order conditions are: j ≠i = λ *j + γ * , j ≠ i (16) * 2 d U ′j ( X ij ) U ′′jU j − U ′j Because we assume concave utility, U ≤ 0 so = < 0 . Hence the left dX ij U j ( X ij* ) U2 '' j hand side of each equation j in the first order condition is monotone (decreasing) and must have an inverse, dubbed I j . So= X ij* I j (λ *j + γ * ) . If X ij* < Lij , the complementary slackness condition implies that λ *j = 0 in (16), in which case the equations yield X ij* = I j (γ * ) ≡ Ci , ∀j:X*ij < Lij . In other words, all creditors of agent i who are not made whole receive equal dollar payments Ci. So X ij* = min( Lij , Ci ) in accord with the CE Rule (15). QED While Proposition 4 shows that the Dagan and Volij (1993) proof can be modified to accommodate heterogeneous concave utilities rather than just their linear utilities, their result equating (liability-) weighted Nash Bargaining with the Proportional Rule does not similarly generalize. To see this, take the log of the weighted objective function to find the Lagrangian for that problem: = l ∑ j ≠i Lij log(U j ( X ij )) − ∑ j ≠i λ j ( X ij − Lij ) − γ (∑ X ij − E ) j ≠i With first order conditions: U ′j ( X ij* ) λ *j + γ * = , j≠i U j ( X ij* ) Lij 17 As before, if X ij* < Lij , λ *j = 0 , and upon inverting the first order conditions one obtains X ij* = I j ( γ* Lij ) . Because I j is monotone decreasing, X ij* is a monotone increasing function of Lij , denoted X ij* = f j ( Lij ) . For some other agent j’, if X ij* ' < Lij ' , f j ' is a possibly different increasing function of Lij ' . This is because it is derived from its own agent’s marginal log utility, and hence may not be the same function. But the Proportional Rule requires that these be the same increasing linear function X ij* = θi* Lij , for j , j ′ . Shummer and Thomson (1997) provided alternative constrained optimization characterizations of the CE Rule. Their Proposition 3 characterizes the CE required payments as the feasible payments minimizing ``the largest amount received by any agent and the smallest such amount” while their Proposition 4 characterizes them as the feasible payments minimizing “the variance of the results received by all the agents”. 10 Due to the close connection between variance and entropy, and the fact that the unconstrained maximum entropy distribution is uniform, the latter characterization suggests the possibility that there is also an entropic characterization of the CE rule. This is indeed the case, as shown below. Proposition 5: Define the (unnormalized) entropy ∑ −X j ≠i ij log X ij of the payments X ij . The CE Rule (15) produces payments that solve the following constrained maximization of entropy: max − ∑ X ij log X ij s.t. j ≠i X ij ≤ Lij , ∀j ≠ i ∑X j ≠i ij (17) = ai Proof: The Lagrangian for the problem (17) is −∑ X ij log X ij + ∑ λ j ( X ij − Lij ) + γ ∑ X ij − ai = j ≠i j ≠i j ≠i The first order conditions result in the following equations for the solutions X ij* : λ* X ij* = e j eγ * −1 . If X ij* < Lij , the complementary slackness condition implies that λ * j = 0 , in = which case the equations yield X ij* eγ * −1 ≡ Ci , ∀j:X*ij < Lij . In other words, all creditors not 10 Readers of these papers should note that my use of the term “feasible” incorporates the constraint that no claimant receive more than it is owed. The propositions in those papers also impose that constraint, but not in their definition of feasible allocations. As such, there is no loss in generality when imposing it in the feasibility condition. 18 made whole receive equal dollar payments. So X ij* = min( Lij , Ci ) in accord with the CE Rule (15). QED Generalizations to Financial Networks: Some Important Differences In a network like our example (1), the systemic consequences of agent #2’s default must also be considered. When implementing the Proportional Rule in Section III, we depended on the existence proof of Eisenberg and Noe (op.cit., p.240) . They created a mapping Φ on the feasible set of payments vectors, whose maximal fixed point determines the Proportional Rule solution. Existence of a maximal (and a minimal) fixed point is implied by the Knaster-Tarski Fixed Point Theorem. While their mapping does not straightforwardly extend to the CE Rule, a different mapping can be used for this purpose. Under the CE rule, the feasible set is defined by a set of nonnegative numbers C1 , …, CN ≡ C such that ∑ min( Lij , Ci ) ≤ ∑ min( L ji , C j ) i = 1, …, N . The left hand j ≠i side is ∑X j ≠i ∑X j ≠i ji ij j ≠i ≡ li (C) , i.e. aggregate payments made by Agent i, while the right hand side is = ai (C) , i.e. the aggregate of payments received by it from the other agents. We prove: Proposition 6: In a network under the CE Rule, the feasible set is nonempty. Moreover, the C * (C1* , …, CN* ) defining its payments X ij* = min( Lij , Ci* ) can be found by solving maximal = max X ij ∑∑ X i j ≠i ij ≡ max C1 ,…,CN ∑∑ min( L , C ) s.t. i j ≠i ij i 1, …, N li (C) ≡ ∑ min( Lij , Ci ) = ai (C) ≡ ∑ min( L ji , C j ), i = j ≠i (18) j ≠i Proof: Using the above notation for aggregate liabilities paid and assets received by the agents, define the vector-valued maps l (C) ≡ (l1 (C), …, lN (C)) and a (C) ≡ (a1 (C), …, aN (C)) on the 1, …, N . This subset is a complete lattice subset S of vectors C in which Ci ∈ [0, max j ≠i Lij ] , i = with the usual ordering ≤ of N-vectors. l (C) is monotone increasing on S , and hence has an inverse. Because a (C) is monotone nondecreasing on S , the map f : S → S ; f (C) = l −1 (a (C)) is monotone on the complete lattice S . A fixed point of f satisfies the constraints in (18). By the Knaster-Tarski Fixed Point Theorem 11 , the map has a set of fixed points which is also a complete lattice (and hence nonempty), and hence has a maximal element. So it can be found by solving (18). QED 2 1 Solving (18) with the liabilities matrix in our example (1), find C = 10, 1 , 5, 3 . The 3 3 payments matrix X ij = min( Lij , Ci ) is shown below: 11 See https://en.wikipedia.org/wiki/Knaster%E2%80%93Tarski_theorem 19 Agent #1 #2 #3 #4 l* 0 0 10 0 10 #1 #2 12 0 12 12 5 3 3 3 #3 5 5 0 5 15 #4 31 0 31 0 62 3 3 3 * a 10 5 15 6 2 36 2 3 3 A* .273 .136 .409 .182 L* .273 .136 .409 .182 (19) Comparing the CE Rule payments (19) to the Proportional Rule payments (6), we see that total liabilities paid were less under the CE Rule (36 2/3 vs. 43.934). The major contributor to this difference is the initially defaulting Agent #2, who owed 70 and paid 10.3 under the Proportional rule, but only 5 under the CE rule. The Inequity Index (10) is the entropy of the distribution of agent liabilities originally owed L = (6.3%, 43.7%,31.3%,18.7%) to the distribution of what is paid: from (19) that is L* = (27.3%,13.6%, 40.9%,18.2%) . Using (10), the Inequity Index I = .340 , substantially higher than its value I = .168 under the Proportional Rule. This is not surprising, because the Inequity Index (10) penalizes deviation of the distribution of percentage liabilities paid from the distribution of original percentages owed. The Proportional Rule resulted in the initially defaulting Agent #2 paying around 24% of total liabilities paid, much closer to the 43.8% of liabilities originally owed than what is required by the CE Rule, which required Agent #2 to pay only about 14% of liabilities paid. This is the major contributor to the increase in the Inequity Index arising from the CE Rule. While Proposition 6 characterizes the extension of the CE Rule to networks, it does not imply that the CE Rule is the Nash Bargaining solution within the network, as it is when there is only a single exogenous defaulter with multiple creditors (see Proposition 4). With a network, the Nash Bargaining solution solves the following problem: max X ij ∏i U i ai ≡ ∑ X ji s.t. j ≠i X ij ≤ Lij li ≡ ∑ X ij = j ≠i ∑X j ≠i ji (20) ≡ ai , i= 1, …, N The Lagrangian is: = ∏ U (∑ i i j ≠i ) X ji − ∑ ∑ λij ( X ij − Lij ) − ∑ γ i ∑ X ij − ∑ X ij i i j ≠i j ≠i j ≠i As in the proof of Proposition 4, we can substitute the log of the product of utilities and write the first order conditions with respect to each X ij as: 20 U ′j ∑ X ji* j ≠i λ * + γ * − γ * , ∀i, j ≠ i = ij i j * U j ∑ X ji j ≠i (21) As in the proof of Proposition 4, for any j we can invert the left hand side to derive ∑ j ≠i X *ji = I j (λij* + γ i* − γ *j ), j ≠ i Now if Agent i pays two agents j and j’ less than they are respectively owed, complementary slackness implies that λij* = λij*′ = 0 . But due to the presence of γ j * and γ j '* in system (21), this does not force X X ij′ ≡ Ci as it does in the first order conditions (16) that lead to Proposition = ij 4. That is, an agent i does not have to pay the same amount to two other agents who each receive less than owed. In networks, the Nash Bargaining solution does not imply the CE rule. To illustrate this using our example (1), we again assume linear utilities and numerically solve (20) to find: Agent #1 # 2 #3 # 4 l* 0 0 10 0 10 #1 #2 0 0 10 20 30 0 30 0 10 40 #3 #4 10 0 20 0 30 * 10 30 40 30 110 a A* .091 .273 .363 .273 L* .091 .273 .363 .273 (22) The Nash Bargaining solution (22) is neither the Proportional Rule solution (6) nor the CE Rule solution (19). Total liabilities paid is higher in the Nash Bargaining solution than in either of them (110 vs. 43.93 and 36 2/3, respectively), and accordingly the Default Severity Index (8) is much lower (.440 vs. 1.68 and 1.81). Moreover, recall that the Inequity Index (10) is the relative entropy of the distribution of agent liabilities owed L = (6.3%, 43.7%,31.3%,18.7%) relative to the distribution of what is paid: from (17) that is L* = (9.1%, 27.3%,36.3%, 27.3%) . The Nash Bargaining solution’s distribution of liabilities paid appears closer to the distribution of what was originally owed than do the distributions resulting from the Proportional Rule L* = (22.8%, 23.5%,39.2%,14.5%) and the CE Rule L* = (27.3%,13.6%, 40.9%,18.2%) . Not surprisingly the relative entropy, i.e. our Inequity Index (10), is only I = .066 with the Nash Bargaining solution, which indeed is lower than the Inequity Index values resulting from the Proportional Rule ( I = .168 ) and the CE Rule ( I = .340 ). We summarize our findings as Proposition 7: 21 Proposition 7: In a network, the Nash Bargaining solution to (20) does not implement either the CE Rule nor the Proportional Rule, even when utilities are linear. At least in our example, it produced both higher efficiency (i.e. higher total liabilities paid or lower Default Severity Index (8)) and lower inequity (10) than those rules did. It is interesting to compare the Nash Bargaining solution with linear utilities (22) to the liabilityweighted Nash Bargaining solution computed from the same liability matrix (1). It is: Agent #1 #2 #3 #4 l* L* 0 0 10 0 10 .091 #1 #2 0 0 13.75 16.25 30 .273 3.75 30 0 10 43.75 .398 #3 #4 6.25 0 20 0 26.25 .278 * 10 30 43.75 26.25 110 a A* .091 .273 .398 .278 (23) In general, the liability-weighting should shift payments toward those who were originally owed more. Comparing (23) to (22), we see that liability-weighting resulted in slightly more going to Agent #3 and slightly less going to Agent #4. This is not surprising, because (1) shows that Agent #3 was owed 50 while Agent #4 was owed only 30. The total liabilities paid are 110, the same as in the (un-weighted) Nash Equilibrium, and the Default Severity Index (8) is .437. However, there are multiple solutions, because (22) achieves the same value of the liabilityweighted Nash objective function as does (23). Not surprisingly, the liability-weighted solution (23) is inconsistent with the Proportional Rule, e.g. Agent #2 pays 0% of the 30 it owed to Agent #1, while paying around 69% of the 20 it owed Agent #3, and around 81% of the 20 it owed Agent #4. In summary, the single-debtor characterization of the CE (Proportional) Rule as a Nash (liability-weighted Nash Bargaining) solution does not generalize to financial networks. An Extension: Permitting Netting Before Resolution Mokal (2015) highlights bankruptcy provisions in many countries that provide priority treatment to counterparties of bankrupt agents who have entered into swaps, repos, and/or some other derivative securities with them. The counterparties are permitted to net their claims against each other before general bankruptcy resolution procedures take place. In our example, suppose the initially bankrupt Agent #2 and is permitted to net its promised cash flows before a bankruptcy resolution procedure. For convenience, the original liabilities matrix (1) is reprinted below: 22 #2 #3 # 4 l Agent #1 #1 0 0 10 0 10 #2 30 0 20 20 70 Lij ≡ #3 10 30 0 10 50 #4 10 0 20 0 30 50 30 50 30 160 a A .313 .187 .313 .187 L .063 .437 .313 .187 Note that Agent #2 owes 20 to Agent #3, who in turn owes 30 to Agent #2. If Agent #2 is permitted to net that liability against what it is owed, Agent#2 will owe nothing to Agent #3, while Agent #3 will owe only 10 to Agent #2: Lnet ij l #2 #3 # 4 Agent #1 #1 0 0 10 0 10 #2 30 0 0 20 50 ≡ #3 10 10 0 10 30 #4 10 0 20 0 30 50 10 30 30 120 a A .313 .187 .313 .187 L .083 .417 .250 .250 (24) Mokal (op.cit.) argues that some policymakers have encouraged such netting prior to bankruptcy proceedings, based in part on the claim that it serves to mitigate contagion. Mokal cites guidance from the U.N. Commission on International Trade Law (2005), which wrote: “Without the ability to close out, net, and set-off obligations…a debtor’s failure to perform its contract…could lead the counterparty to be unable to perform its financial contracts with other market participants. The insolvency of a significant market participant could result in a series of back-to-back transactions, potentially causing financial distress to other market participants, and in the worst case, resulting in the financial collapse of other counterparties, including regulated financial institutions. This domino effect is often referred to as systemic risk, and is cited as a reason as a significant policy reason for permitting participants to close out, net, and set off obligations in a way that normally would not be permitted by insolvency law.” While Mokal (op.cit.) persuasively argues that this source of contagion was not a significant risk factor in the Financial Crisis of 2008 12, it is interesting to study its effects within the context of our example, meant to model a relatively severe contagion situation that could arise in payments and trading networks. 12 See Mokal (op.cit., p. 45). In fact, Mokal (op.cit.) further argues that an alternative contagion channel (precipitated by adverse common shocks, resulting in asset sale-induced decreases in collateral values and subsequent bankruptcies induced by inadequate collateralization) did play an important role in the Financial Crisis, and that netting exacerbates that contagion channel. Hence Mokal opposes netting prior to bankruptcy resolution. 23 Substituting (24) for (5) will affect the bankruptcy resolution payments matrix X ij . If the Proportional Rule is used, substitute (24) into (5) and solve to find the clearing fractions paid θ*net = (100%,9.7%, 48.4%, 22.6%) vs. the previously calculated fractions θ* = (100%, 14.8%, 34.4%, 21.3%) . Instead of the resolution matrix (6), netting prior to bankruptcy will produce the resolution matrix: Agent #1 #2 X ij ≡ #3 #4 * a A* L* .277 .134 .402 2.258 0 4.516 0 6.774 .1875 10 4.839 14.516 6.774 36.129 .277 .134 .402 .188 #1 #2 0 0 2.903 0 4.839 4.839 #3 10 0 0 #4 l* 0 10 1.935 4.839 4.839 14.516 (25) So in this example, permitting the bankruptcy triggering Agent #2 to net its exposures prior to resolution does not deter contagion – Agents #2, #3, and #4 still wind up defaulting. In fact, netting before applying the Proportional Rule results in a slight increase in the Default Severity Index S from 1.46 (calculated using the original liabilities matrix (1) and the Proportional Rule’s resolution matrix (6)) to 1.53 (calculated using the netted liabilities matrix (24) and the Proportional Rule’s resolution matrix (25)), and an increase in the Inequity Index I from 0.168 to 0.326. Of course, the Nash Bargaining rule analyzed earlier will also result in a different resolution payments matrix when netting is permitted. In our example, the Nash Bargaining solution yields: Agent #1 # 2 #3 # 4 0 0 10 0 #1 #2 0 0 0 10 X ij ≡ #3 10 10 0 10 #4 0 0 20 0 * 10 10 30 20 a A* .143 .143 .428 .286 l* 10 10 30 20 70 L* .143 .143 .428 .286 (26) Netting before applying the Nash Bargaining Rule results in an increase in the Default Severity Index S from 0.440 (calculated using the original liabilities matrix (1) and the Nash Bargaining resolution matrix (22)) to 0.772 (calculated using the netted liabilities matrix (24) and the Nash Bargaining resolution matrix (26)), and an increase in the Inequity Index I from 0.066 to 0.233. Hence in this example, netting does not improve default severity or inequity of its resolution. 24 VIII. Conclusions The mutual information statistic is defined using the Kullback-Leibler relative entropy. It has heretofore been used as an objective function in constrained minimization problems for estimating unknown cells in inter-agent payments matrices, and analogous matrices arising in the social sciences. Inter-agent payments matrices are inputs to studies estimating the frequency and severity of default, which includes both initial defaults and secondary defaults induced by the initial defaults (a.k.a. contagion). The mutual information is a measure of dependence between the distribution of agents’ respective liabilities (money owed) and assets (money that is owed to them), but it is positive regardless of whether that dependence is positive or negative. It is argued that only negative dependence (e.g. negative Kendall rank correlation) clearly results in more default and contagion severity. So it isn’t surprising that the simulations conducted in this paper did not demonstrate a fixed relationship between the mutual information in the liabilities matrix and measures of default and contagion severity, including new, relative entropy-based measures devised for and used throughout this paper. The standard default resolution procedure forces a defaulting party to pay the same percentage of what it owes each of its claimants. For example, if a defaulting party has enough funds to pay 80% of its total claims, the standard procedure requires it to pay 80% of what it owes to each of its claimants. This is dubbed the Proportional Rule. When a single agent’s default triggers other defaults (i.e. there is contagion), it is shown that relaxing this restriction it will often result in both lower default and contagion severity as well as less inequity in recovery. Minimizing inequity rather than default severity is a feasible alternative objective, but may not be nearly as important as relaxing the strict proportional payments constraint on agents that trigger defaults. In light of these findings, investigation of the game-theoretic bargaining literature uncovers reasons favoring alternative default resolution rules. But that literature has heretofore only considered the case where only one agent is a debtor to multiple claimants. In that limited setting, the Nash Bargaining solution requires equal dollar payments (when not greater than what is owed), rather than percentage payments. This is termed the Constrained Equality (CE) Rule. Herein, the game-theoretic analysis was extended to financial networks, where potentially all agents are both debtors and claimants. A fixed point argument was used to prove that there will be a CE Rule resolution payments matrix, but it was also shown the Nash Bargaining solution no longer implies that it be used. In fact, the Nash Bargaining solution will not feature either fixed percentage or fixed dollar payments. The same negative result is obtained when comparing the Proportional Rule to the liability-weighted Nash Bargaining solution. A computed example shows that the Nash solutions can dominate both of these rules, in terms of less default severity and less inequity in bankruptcy resolution payments. 25 REFERENCES Demange, G. (2015). Contagion in financial networks: A Threat Index (CESifo Working Paper No. 5307). Retrieved from www.CESifo-group.org/wp Dagan, N. and Volij, O. The bankruptcy problem: a cooperative bargaining approach. Mathematical Social Sciences, 26, 287–297. Eisenberg, L. and Noe, T. (2001). Systemic risk in financial systems. Management Science, 47(2), 236-249. Elimam, A., Girgis, M., and Kotab, S. (1996). The use of linear programming in disentangling the bankruptcies of Al-Manakh stock market crash. Operations Research, 44(5), 665-676. Kadens, E. (2010). The last bankrupt hanged: balancing incentives in the development of bankruptcy law. Duke Law Journal, 59(7), 1229-1319. Levinthal, L. (1918). The early history of bankruptcy law. University of Pennsylvania Law Review, 66, 223-250. Mokal, R. J. (2015). Liquidity, systemic risk, and the bankruptcy treatment of financial contracts. Brooklyn Journal of Corporate, Financial, and Commercial Law (forthcoming). Rogers, L.C.G. and Veraart, L.A.M. (2013). Failure and rescue in an interbank network. Management Science, 59(4), 882-898. Sachs, A. (2014). Completeness, interconnectedness and distribution of interbank exposures—a parameterized analysis of the stability of financial networks. Quantitative Finance, 14(9), 16771692. Thomson, W. (2003). Axiomatic and game-theoretic analysis of bankruptcy and taxation problems: a survey. Mathematical Social Science, 45, 249-297. Thomson, W. (2015). Axiomatic and game-theoretic analysis of bankruptcy and taxation problems: an update. Mathematical Social Science, 74, 41-59. Upper, C. (2011). Simulation methods to assess the danger of contagion in interbank markets. Journal of Financial Stability, 7, 111-125. Upper, C. and Worms, A. (2004). Estimating bilateral exposures in the German interbank market: Is there a danger of contagion? European Economic Review, 48, 827-849. 26 FIGURES 27 28