Mortgage Burnout and Selection Effects in Heterogeneous Cox Hazard Models

1. Problem Statement

The paper addresses the phenomenon of mortgage burnout, a well-documented empirical feature in mortgage-backed securities (MBS) where the observed prepayment rate of a mortgage pool declines over time (seasoning), even when the aggregate refinancing incentive (e.g., interest rate differentials) remains favorable.

• The Puzzle: Standard models often struggle to explain why prepayment rates drop as a pool ages without invoking complex behavioral changes in borrower decision-making.
• The Hypothesis: The author posits that burnout is not necessarily a behavioral shift but a statistical selection effect arising from borrower heterogeneity. In a heterogeneous population, borrowers with the highest propensity to refinance (high hazard rates) exit the pool first. Consequently, the surviving population becomes progressively "enriched" with low-hazard borrowers, causing the aggregate hazard rate to decline naturally.
• The Gap: While this mechanism is understood in deterministic hazard models, there is a lack of rigorous mathematical formulation for stochastic intensity models (specifically Cox processes), which are more realistic for modeling mortgage prepayments where hazard rates fluctuate due to common market factors.

2. Methodology

The paper utilizes a measure-theoretic approach within a filtered probability space to model a heterogeneous population of mortgages.

• Model Setup:

  • Let $x$ represent a borrower type in a type space $\mathcal{X}$.
  • Each borrower has an individual hazard rate $\lambda_t(x)$, modeled as a non-negative adapted intensity process (Cox process).
  • The survival process for a borrower of type $x$ is $S_t(x) = \exp(-\int_0^t \lambda_s(x) ds)$.
  • The aggregate pool survival is the weighted average: $\bar{S}_t = \int S_t(x) F(dx)$, where $F$ is the initial distribution of borrower types.

• Key Definitions:

  • Observed Pool Hazard ($\bar{\lambda}_t$): Defined as $-\frac{d}{dt} \log \bar{S}_t$.
  • Survival-Weighted Measure ($F_t$): A time-dependent probability measure defined as $F_t(dx) = \frac{S_t(x)F(dx)}{\bar{S}_t}$. This represents the distribution of borrower types among those who have not yet prepaid.
  • Expectations: $E_t[\phi]$ denotes the expectation of a function $\phi$ under the survival-weighted measure $F_t$.

• Derivation Strategy:
  The author derives the dynamics of the pool hazard by differentiating the relationship $\bar{\lambda}_t = E_t[\lambda_t(x)]$. This involves applying Itô calculus (for stochastic cases) and standard differentiation (for deterministic cases) to the ratio of integrals defining the expectation under the changing measure $F_t$.

3. Key Contributions

The paper's primary contribution is the formulation of a general stochastic identity that decomposes the evolution of the aggregate pool hazard into distinct structural components.

A. The General Identity

The paper establishes that the change in the observed pool hazard is driven by three factors:

1. Mean Drift: The average evolution of individual hazards.
2. Selection Effect (Variance Term): A negative term proportional to the cross-sectional variance of hazards.
3. Diffusion Term: A stochastic term inherited from common market factors.

B. Connection to Evolutionary Biology

The author draws a profound conceptual parallel between mortgage burnout and the Price Equation from evolutionary biology.

• In biology, the Price equation decomposes the change in a trait's mean into a selection component (covariance between fitness and the trait) and a transmission component.
• In this context, "fitness" is survival (remaining in the pool), and the "trait" is the hazard rate.
• The paper demonstrates that mortgage burnout is mathematically equivalent to natural selection: high-hazard borrowers are "selected out" (prepay), leaving a population with a lower mean hazard.

4. Key Results

4.1 Deterministic Hazard Case

If individual hazards $\lambda(t, x)$ are deterministic and differentiable, the pool hazard evolves as:
$$\frac{d}{dt}\bar{\lambda}(t) = E_t[\dot{\lambda}(t, x)] - \text{Var}_t(\lambda(t, x))$$

• Result: Even if individual hazards are constant ($\dot{\lambda}=0$), the pool hazard declines monotonically ($\frac{d}{dt}\bar{\lambda} = -\text{Var}_t < 0$) solely due to heterogeneity.

4.2 Stochastic Hazard Case (Cox Processes)

When individual hazards follow an Itô process $d\lambda_t(x) = \mu_t(x)dt + \sigma_t(x)dW_t$ (driven by a common factor $W_t$), the pool hazard evolves as:
$$d\bar{\lambda}_t = \left( E_t[\mu_t(x)] - \text{Var}_t(\lambda_t(x)) \right) dt + E_t[\sigma_t(x)] dW_t$$

• Drift: Contains the mean drift of individual hazards minus the cross-sectional variance (the burnout term).
• Diffusion: The volatility of the pool hazard is the survival-weighted average of individual volatilities.
• Significance: Heterogeneity induces a systematic downward drift in the observed hazard, even when individual hazards are stochastic. This confirms that burnout is a structural feature of heterogeneous Cox populations.

4.3 Measure Change Interpretation

The paper interprets the survival-weighted measure $F_t$ as a Radon-Nikodym tilt of the initial distribution $F$. The weighting factor is proportional to survival probability. This shows that the observed decline in prepayment rates is a result of the cross-sectional distribution of borrower types becoming endogenously tilted toward lower-hazard types over time.

4.4 Frailty Factor Models

The author applies the general theory to specific frailty models commonly used in MBS:

• Gamma Frailty: With a constant common factor, the model yields a hyperbolic decay of the pool hazard ($\bar{\lambda}_t \propto \frac{1}{1 + \theta \lambda t}$). This is observationally equivalent to a deterministic model with time-varying hazard.
• Lognormal Frailty: For small dispersion, the burnout profile is approximately exponential.
• Normal Frailty: For small variance, the burnout effect is approximately linear over short horizons.

5. Significance and Implications

1. Structural Explanation of Burnout: The paper provides a rigorous, model-free explanation for mortgage burnout. It proves that burnout arises generically in any heterogeneous population where high-hazard individuals exit first, without requiring assumptions about changing borrower behavior or "learning."
2. Stochastic Generalization: By extending the deterministic burnout identity to Cox processes, the paper bridges the gap between simple survival analysis and complex stochastic intensity modeling used in modern quantitative finance.
3. Modeling Efficiency: The results suggest that complex behavioral models might be unnecessary to capture the "burnout" phenomenon; instead, it can be captured structurally through the variance term in the hazard dynamics.
4. Universal Mechanism: By linking the phenomenon to the Price Equation, the paper elevates mortgage burnout from a specific financial anomaly to a universal statistical selection mechanism applicable to any heterogeneous population subject to exit risks.

In summary, Lesniewski demonstrates that the "burnout" of mortgage pools is a mathematical inevitability of heterogeneity, driven by a negative variance term that acts as a selection filter, systematically lowering the average hazard rate of the surviving population.