The manuscript "Flood risks to the financial stability of residential mortgage borrowers: An integrated modeling approach" presents an impressive framework for evaluating how flooding impacts mortgage borrowers' financial stability. The authors develop a comprehensive modeling approach that integrates flood damage estimates, property values, mortgage balances, and insurance coverage to identify which borrowers face heightened exposure to mortgage default following flood events.

The study makes a valuable contribution by introducing a model framework that models the relationship between pre-flood financial conditions, flood impacts, and post-flood financial conditions. The integrated modeling approach allows for the assessment of different theorized causal pathways of mortgage default (strategic, cashflow, and double-trigger).

While I appreciate many aspects of the study, I believe it faces several limiting issues that I would like to see the authors address in a revision:

1. Unclear framing: The study lacks a clearly articulated research question that aligns with its methodological approach. While the introduction suggests the study addresses "how pre-flood financial conditions affect the relationship between uninsured damage exposure and post-flood risk of mortgage default," the methods and results don't directly answer this question. The study would benefit from explicitly stating what scientific questions it aims to answer and how its modeling approach addresses these questions.
2. Absence of calibration/validation data for defaults: Despite citing empirical research on drivers of default, the authors don't calibrate or validate their framework against observed mortgage outcomes. Without these crucial modeling steps, it's difficult to assess whether the modeled financial conditions offer predictive value. The utility of the integrated modeling framework is questionable without demonstrating its predictive accuracy for the outcome of interest, unless the authors pursue a more exploratory approach with more detailed uncertainty quantification and sensitivity analysis.
3. Unconvincing causal mechanism of flood damage to default: In light of the comment above, the findings from Kousky et al. (2020), a key reference for the authors, demonstrate significant concerns with using modeled damage estimates to predict mortgage outcomes. Their study shows that catastrophe model damage estimates—even those potentially more accurate than those in the current study because they come from a proprietary catastrophe model—found spurious relationships between predicted flood damage and default compared to results based on actual damage inspections. Specifically, they found that for rare events like default, "predicted damage needs to match better with actual damage at a property level in order to deliver a robust estimated impact." As another example, they found that when the catastrophe model predicted damage of less than 10%, the odds of deep delinquency or default increased, but not when the catastrophe model predicted greater than 10% damage. They wrote, “This counter-intuitive risk ranking, which we have not seen in other loan performance outcomes, suggests that the inaccurate property-level damage prediction by the catastrophe model can be problematic for a rare outcome, such as deep delinquency or default.” This raises fundamental questions about the reliability of the current study's approach to modeling default risk.
4. Potential need for reframing study around sensitivity analysis: Despite the complex integrated modeling approach, the paper doesn't sufficiently explore how uncertainties in model components propagate through to default projections. A more comprehensive uncertainty quantification and sensitivity analysis would strengthen the study by identifying which factors most influence projected outcomes and how robust the findings are to different assumptions. This approach would be particularly valuable given the lack of validation data and would better demonstrate the framework's utility for policy analysis.
5. Missed opportunity for policy analysis: The study introduces interesting policy analyses (such as the home repair grant program) but doesn't fully leverage its framework to explore how various policies could influence default rates under different scenarios and assumptions. A more thorough exploration of policy interventions, coupled with comprehensive sensitivity analysis, would significantly enhance the paper's contributions and better justify the development of the integrated modeling approach.
6. Omission of important contextual factors: The study takes a real-world framing, which is compelling and raises the stakes of the findings, but the model excludes real-world factors that would influence default outcomes, such as disaster aid programs, employment changes, and the effects of the 2008 financial crisis, without adequate justification for these simplifications.

Please see the attached file for my full review, which includes detailed comments that expand on these high-level synthesis points for each section of the paper.

Review of “Flood risks to the financial stability of residential mortgage borrowers: An integrated modeling approach”

General Comment Statement

The paper presents a significant and well documented contribution to the field of climate-financial risk. The integrated, "bottom-up" modelling framework, which links property-level flood damage to household financial distress, is a commendable and ambitious effort to advance the understanding of this critical issue.

As I am not an expert in US insurance market, I focused the review mainly on the modelling part of the work. While the framework is conceptually sound, the analysis concludes that its current implementation contains a cascade of methodological limitations that is likely going to lead to a systematic underestimation of the true financial risk.

Line 18: The finding that the evaluated floods "generated $4.0 billion in property damage" is a key quantitative output. However, this figure should be interpreted as a conservative bound. As mentioned below (see comments on Line 271), the damage detection model fails to identify a majority of properties that actually sustained damage, meaning the true total damage might be higher.

Line 20: The statement that 32% of affected borrowers lacked sufficient income or collateral, "placing them at an elevated risk of default," is based on the underwriting criteria used. Is there any risk that the criteria used could lead to an underestimation of the number of borrowers who would be denied credit and thus be at risk?

Lines 234-240: The generation of "pseudo-absence" points is a pragmatic solution to incomplete data but introduces noise. The authors' own validation (Line 276) shows that model precision increases significantly when these points are excluded, suggesting that a number of these randomly generated "undamaged" points likely distorted the model's training (actually damaged?).

Line 271: A very low sensitivity of just 12% to 42% means the model fails to identify between 58% and 88% of properties that were actually damaged. This is a foundational error that guarantees a systematic underestimation of the total number of impacted households and the total damage costs. All subsequent risk estimates are therefore performed on a small fraction of the true at-risk population.

Lines 272-276: The authors' framing of this result is misleading. The model's high precision is emphasized while downplaying the severe consequence of the high false-negative rate. In risk assessment, particularly for disaster aid, the cost of a false negative (failing to identify a household in need) is high.

Lines 343-345: The reported accuracy is a major concern. Only 54% of the model's value predictions fall within ±20% of the actual sale price. The authors later note this is the largest source of uncertainty in their final results (Lines 674).

Lines 357-360: Is the use of GSE data to model the entire market, creating a bias of the “typical” borrowing population? If yes, it should be stipulated to keep the modelling results in perspective.

Line 583: The conclusion must be interpreted as a conservative floor, not a central estimate, due to the cascading methodological issues outlined above and should be mentioned as such.

Line 494-675: While the paper's focus is on flood risk, its analysis spans a period in which the U.S. housing market underwent its most significant shock in generations. From 2008, the model might overlook a critical variable that shaped housing values, credit availability, and the underlying financial health of borrowers.