Contagion or Macroeconomic Fluctuations? Identifiability in Aggregated Default Data

This paper investigates the identifiability of contagion in aggregated default data by comparing different dependence structures, finding that while macroeconomic fluctuations generally dominate annual default clustering, a persistent contagion signal remains distinguishable in the Lo-Davis model but is effectively absorbed into macroeconomic variation in the Torri model.

The Gist

The Big Question: Is it a Virus or the Weather?

Imagine you are watching a crowd of people in a stadium. Suddenly, a bunch of them start leaving their seats at the same time. You want to know why.

There are two main theories:

1. The "Virus" Theory (Contagion): One person sneezed, and the people next to them caught a cold, who then sneezed on others. The leaving is spreading from person to person like a chain reaction.
2. The "Weather" Theory (Macroeconomics): A sudden storm cloud appeared over the whole stadium. Everyone got wet at the same time and decided to leave because of the rain, not because of each other.

The Problem: You are standing far away in the nosebleed section. You can only see the total number of people leaving each hour. You can't see who sneezed on whom, and you can't see the rain clouds clearly. You just see the numbers go up and down.

The Paper's Goal: The author, Shintaro Mori, asks: Can we look at the total numbers and figure out if it's a "virus" spreading or just "bad weather" affecting everyone?

---

The Three Models (The Tools)

To solve this, the author tests three different "mathematical lenses" to see which one fits the real-world data best. Think of these as three different ways to explain the crowd leaving:

1. The "Smooth Rain" Model (Vasicek Model):

   • How it works: Imagine the weather changes gradually. A little rain makes a few people leave; a heavy storm makes many leave. The probability of leaving is tied to a single, smooth "weather factor" that affects everyone equally.
   • The Analogy: It's like a dimmer switch for the lights. You can turn it up or down smoothly, and the room gets brighter or darker gradually.

2. The "Domino" Model (Lo–Davis Model):

   • How it works: This is a cumulative virus. If one person leaves, they make the person next to them slightly more likely to leave. If two people leave, the pressure on the third person gets even higher. The more people who leave, the faster the rest follow.
   • The Analogy: It's like a game of dominoes. Knocking over one makes the next one fall, which makes the next one fall harder. The effect builds up smoothly as more dominoes fall.

3. The "Switch" Model (Torri Model):

   • How it works: This is a threshold virus. Nothing happens until one specific person (a "super-spreader") leaves. Once that one person leaves, a switch flips, and suddenly everyone who isn't immune runs for the exit at once. It's an "all-or-nothing" event.
   • The Analogy: It's like a dam breaking. The water level rises slowly (nothing happens), but once it hits a specific line, the dam bursts, and everything floods out instantly.

---

The Experiment: What Happened?

The author looked at 100 years of real data (1920–2023) regarding companies going bankrupt (defaults). He compared the three models against the actual numbers.

Phase 1: The "Steady State" Test (Ignoring Time)

First, he pretended the economy never changed. He just looked at the numbers as if they were random snapshots.

• Result: The "Smooth Rain" (Vasicek) model won easily. It fit the data best, especially for rare, huge crashes.
• Why? The "Smooth Rain" model is very flexible. It can mimic almost any pattern of clustering just by adjusting the "weather." The "Switch" model was too rigid and couldn't explain the data well.

Phase 2: The "Real World" Test (Adding Time)

Then, the author realized the economy does change. Some years are great (sunny), some are terrible (storms). He added a "time" variable to the models, allowing the "base probability" of default to change from year to year.

• The Big Discovery: Once he accounted for the changing economy (the weather), most of the "clustering" disappeared.
  • It turns out that when companies fail in groups, it's mostly because the economy got bad, not because one company's failure caused another's.
  • The "Virus" (contagion) is actually a much smaller player than we thought.

Phase 3: Who Can Still Be Seen?

Here is where it gets interesting. Even after accounting for the "weather," the author looked for the tiny leftover "virus" signal.

• The "Switch" Model (Torri): The virus signal vanished completely. It was impossible to tell if a "switch" had flipped or if it was just the weather. The "Switch" model's effect got completely swallowed up by the macroeconomic changes.
  • Analogy: If a storm is raging, you can't tell if a single person sneezed and started a chain reaction. The noise of the storm drowns out the sneeze.
• The "Domino" Model (Lo–Davis): A tiny, persistent signal remained. Even after accounting for the bad economy, this model showed a small, distinct pattern that looked like a virus.
  • Analogy: Even in a storm, if you look closely, you can still see a specific ripple pattern that suggests a domino effect is happening, just a very weak one.

---

The Takeaway: What Does This Mean?

1. Macro is King: When we see a bunch of companies failing at once, it is almost always because the whole economy is in trouble (the "Weather"), not because of a chain reaction of failures (the "Virus").
2. Data is Blurry: If you only look at yearly totals (aggregated data), it is very hard to prove that a "virus" exists. The "Switch" type of contagion is invisible in this data; it looks exactly like bad weather.
3. The Exception: The only type of contagion that leaves a tiny, visible fingerprint in the yearly numbers is the "Domino" style (cumulative), where the risk builds up slowly. But even that is a small effect compared to the economy.

In Simple Terms:
If you see a crowd running out of a stadium, don't immediately assume someone started a panic. It's much more likely that the roof is leaking (the economy is bad). If you want to know if there was a panic, you need to look at the exact second-by-second movement of people, not just the total count at the end of the hour. With just the total count, the "panic" (contagion) is usually hidden by the "leaking roof" (macroeconomics).

Technical Summary

1. Problem Statement

The central problem addressed is the identifiability of contagion in aggregated credit risk data. In credit risk modeling, default clustering (multiple firms defaulting simultaneously) is observed empirically. Two primary mechanisms are proposed to explain this:

1. Contagion: A micro-level mechanism where the default of one firm directly increases the probability of default for others (e.g., through financial networks).
2. Macroeconomic Fluctuations: A common shock (systemic factor) that simultaneously increases default probabilities across all firms.

The challenge arises because empirical data is often limited to aggregated annual default counts (e.g., total defaults per year) rather than firm-level event times or network linkages. The paper asks: Can we distinguish between contagion and macroeconomic fluctuations when observing only coarse-grained, aggregated data? Specifically, does contagion generate a unique statistical signature in the distribution of default counts that remains distinct from the effects of time-varying macroeconomic conditions?

2. Methodology

The author employs a comparative modeling approach using three distinct dependence structures, all of which can be viewed as binomial-mixture models but with different underlying mechanisms.

A. The Models

1. Lo–Davis Model (Cumulative Contagion):

   • Mechanism: Contagion is cumulative. The probability of a surviving firm defaulting increases smoothly as the number of prior defaults ($k$) increases.
   • Structure: $Z_i = X_i + (1-X_i)(1 - \prod_{j \neq i} (1 - Y_{ij}X_j))$.
   • Characteristic: Generates a smooth, $k$-dependent contagion mechanism.

2. Torri Model (Threshold-Type Contagion):

   • Mechanism: Contagion is triggered by a global "OR" condition. If at least one infectious default occurs, a global contagion state is activated, exposing all non-immune firms simultaneously.
   • Structure: Depends on a global indicator $I_n^C = \mathbb{1}(\sum X_j V_j > 0)$.
   • Characteristic: A discrete regime-switching mechanism (non-contagious vs. contagious states).

3. Vasicek Model (Common-Factor Dependence):

   • Mechanism: Defaults are conditionally independent given a continuous latent macroeconomic factor ($F$).
   • Structure: $Z_i = \mathbb{1}(\sqrt{\rho_A}F + \sqrt{1-\rho_A}\epsilon_i \leq \Phi^{-1}(p))$.
   • Characteristic: Generates a smooth mixture of binomial distributions induced by a continuous factor.

B. Empirical Strategy

The study utilizes annual default count data from 1920–2023, categorized into Speculative Grade (SG), Investment Grade (IG), and Aggregate (ALL). The analysis proceeds in two stages:

1. i.i.d. Specification: Assumes constant default probabilities across years. The models are fitted to the data to compare likelihoods and tail behaviors (Value-at-Risk and Expected Shortfall) under the assumption that all variation is due to within-year dependence.
2. Hierarchical Extension: Introduces time-varying default probabilities ($p_t$) via a latent factor to account for cross-year macroeconomic heterogeneity.
   • Variance Decomposition: The total variance of annual default counts is decomposed into:
     • Idiosyncratic variation (i.i.d. component).
     • Dependence/Contagion component (within-year).
     • Time-variation component (cross-year heterogeneity in $p_t$).

3. Key Contributions

• Identifiability Framework: The paper formalizes the question of whether contagion leaves a "structural signature" in aggregated data that cannot be absorbed by macroeconomic variation.
• Structural Comparison: It provides a rigorous comparison of how cumulative contagion (Lo–Davis), threshold contagion (Torri), and common factors (Vasicek) manifest in loss distributions, particularly in the tails.
• Decomposition of Variance: By introducing hierarchical models, the study separates the contribution of time-varying credit conditions from within-year contagion, offering a clearer view of what drives default clustering in aggregate data.

4. Key Results

A. i.i.d. Specification (Constant Parameters)

• Best Fit: The Vasicek model provides the best overall fit (lowest negative log-likelihood) and most accurately reproduces the empirical tail behavior.
• Tail Behavior:
  • The Vasicek model produces a smooth, gradual tail decay matching the data.
  • The Lo–Davis model produces a smoother distribution than Torri but still underestimates extreme events compared to Vasicek.
  • The Torri model exhibits rapid tail decay (suppression of large events) or, depending on parameters, highly volatile tail risks, failing to match the empirical distribution consistently.
• Implication: Under a static assumption, a continuous mixture mechanism (Vasicek) captures default clustering better than discrete contagion mechanisms.

B. Hierarchical Extension (Time-Varying Parameters)

When allowing the default probability $p$ to vary across years (simulating macroeconomic shifts):

• Dominant Driver: The fluctuation of the default probability ($p_t$) across years explains the vast majority of the variance in annual default counts.
• Residual Contagion:
  • Torri Model: Once cross-year heterogeneity is accounted for, the threshold-type contagion component becomes statistically indistinguishable from zero. The model effectively collapses to a purely heterogeneous model; the "contagion" signal is absorbed into macroeconomic variation.
  • Lo–Davis Model: A small but persistent contagion component remains visible in both variance decomposition and tail behavior. Cumulative contagion leaves a detectable structural signature even after aggregation and accounting for macro shocks.
• Tail Risk: The hierarchical Lo–Davis model provides the most balanced fit to empirical tail risk (VaR and ES), whereas the hierarchical Torri and Vasicek models tend to overestimate tail risk (producing excessively heavy tails).

5. Significance and Conclusion

The paper concludes that default clustering in aggregated annual data is driven primarily by time-varying credit conditions (macroeconomic fluctuations) rather than strong within-year contagion effects.

• Identifiability Limits: In coarse-grained data, threshold-type contagion (Torri) is effectively unidentifiable because its effects are indistinguishable from macroeconomic heterogeneity.
• Residual Signal: Cumulative contagion (Lo–Davis) is identifiable to a limited extent, leaving a small but persistent component in the variance and tail distribution that cannot be explained solely by macro shocks.
• Practical Implication: For risk managers and regulators relying on aggregate data, assuming a common-factor structure (like Vasicek) or a cumulative contagion structure is more robust than assuming threshold-type contagion. However, the primary source of risk variation is the evolution of the macroeconomic environment, not the propagation of defaults within a single year.

The study suggests that to truly identify micro-level contagion pathways, higher-frequency data or firm-level network information is required, as aggregation acts as a coarse-graining procedure that suppresses discrete interaction effects.