1. Introduction

There has been much discussion about the effects of climate change and extreme weather on banks’ loan portfolios, and there seems to be a strong interest among both regulators and bankers to develop stress tests around such events. Moreover, modelers of expected losses, like CECL, should consider the implications of these extraordinary events on statistical relationships and historical averages of components of their model suites, and might wish to control for them.

Given that regulatory and industry-wide interest and our own curiosity, we investigated three of the most costly and destructive recent hurricanes to see if default rates, losses given default (LGDs), and overall loss rates were higher in each storm’s path than in both (1) non-path locations and (2) during hurricane-free times. We focused on events that mainly affected one state; so, we ignored destructive storms like Sandy in 2012.

We analyzed three hurricanes:

1. Harvey in Texas in 2017,
2. Irma in Florida in 2017, and
3. Katrina in Louisiana and Mississippi in 2005.

While analyzing Katrina, we noticed an increase in defaults in Louisiana (LA) in 2016/17 and were able to attribute that to flooding from an unnamed storm; so, we analyzed that fourth event in a similar fashion.

For this exercise, we couldn’t use banks’ confidential portfolio data; so, we used Fannie Mae’s publicly available mortgage loan data, which provided large sample sets. However, we caution the reader about projecting our findings to bank-held mortgages or other types of loans.¹

For each event, we determined the hurricane’s path and the duration of its influence, in quarters, and separated contemporaneous loans during its influence into path and non-path portfolios; and grouped loans in hurricane-free times in their own portfolio. We then compared:

1. Default rates across time,
2. Default rates across categories,
3. LGDs across categories, and
4. Overall loss rates across categories.

For B., C., and D., the comparisons were made on a risk-adjusted basis.

As many banks do, we defined “default” as ≥ 90 days past due. Hurricanes Harvey and Katrina and the flooding occurred in late August and Irma in early September; therefore, we wouldn’t anticipate seeing higher defaults or losses in the third quarter. We chose the duration of the hurricane’s effect as consecutive quarters with abnormally high defaults after the storm.²

---

2. Hurricane Harvey, Texas, 2017

Harvey hit the Houston area during August 25 – 29, 2017.

2.A. Harvey’s Path

Tropical cyclone damage is generally driven by flooding; so, we used the USGS post-Harvey flooding report to classify Texas counties as in (red) or outside (blue) of Harvey’s path.³

2.B. Default Rates in Texas through Time

We selected three quarters of abnormally high default rates and compared those rates for each of the three circumstances: (a) within the path, (b) outside of the path, and (c) outside of the hurricane’s period of influence, which we call “hurricane free times.”

Notice the huge, red spike in defaults for loans in the path compared to the blue rates outside of the path. The blue rates are similar to those in adjoining hurricane free times. However, this comparison is not complete because it does not control for risk.

We could measure risk on multiple dimensions, but to keep things simple and clear, we chose a robust, intuitive, theoretically-based, and sound measure that we call the “Spero Portfolio Risk Measure.” It’s constructed on a loan-by-loan basis and aggregated to get overall portfolio or sub-portfolio measures. We use similar measures in loss forecasting modeling, and it–along with other innovations–allows us to use one model for both CECL and stress tests, without any compromise. It ranges from (-∞,∞), where -∞ means risk-free, and +∞ means doom. For this exercise, we did not include other factors like FICO scores, but we compared FICO scores across our three categories and saw little difference.⁴

2.C. Comparison of Risk-adjusted Default Rates

For each of the three categories, we grouped loans by our Spero Portfolio Risk Measure and calculated average default rates. We can’t show the ~37 million observations separately, and if we did, there would be no rates−just a lot of non-defaulted zeroes and a one (or 100%) for each defaulted loan observation. As we note on each graph, the bin sizes vary by category, e.g., the green dots represent ~35 million of the ~37 million observations for Texas; so, each green dot represents many more loans that each blue dot, which represents many more loans than each red dot. Note also that we are ignoring time; we are grouping observations solely on the basis of risk measures, and, therefore, observations from different time periods will be grouped together.

Notice that quarterly default rates within the storm’s path are substantially higher than those outside the path, while those outside of the path are similar to default rates in hurricane-free times.

Geez, that makes the storm look very bad, but what about LGDs?

2.D. Comparison of Risk-adjusted LGDs

LGDs show almost the reverse pattern, but the differences aren’t as great as with default rates.

LGDs during Harvey are similar whether they’re inside or outside the path, but both are uniformly lower than hurricane-free LGDs. So, where Harvey-related default rates seem to be three-to-five times higher than the other groups, LGDs in normal times seem to be about twice Harvey-related LGDs. We’d argue that the difference with hurricane-related LGDs could be related to receiving insurance claims or FEMA assistance or might be strategic defaults or delays in payment because delays were permitted.

2.E. Comparison of Risk-adjusted Charge-off Rates

So, when default rates and LGDs are multiplied together, what happens to charge-off rates? The average charge-off for loans in the path seems to be about two basis points (2 BPs) higher than in hurricane free times.

And that small difference could justify a tiny CECL overlay for a few quarters after a storm.

It is worth noting that by our measure of risk, the riskiest observations arose in relatively normal weather conditions−the right side of the graph and do correspond to the highest observed default rates on the vertical axis.

Another way to visualize net charge-offs is through the interactive heatmap below, which reinforces the conclusions we came to from our scatterplot above. Note that observation counts can vary from tile to tile, and ranges with less than 10k observations are filtered out.

 

3. Hurricane Irma, Florida, 2017

3.B. Default Rates in Florida through Time

As with Texas, we plot default rates by time and category. However, unlike Harvey, here, we see higher default rates outside of the path. Maybe we should have selected more counties as being within the path, but what’s done is done, and the analysis proceeds. Those in our path comprise 40% of all observations in Florida during Irma, while observations in our Harvey path only contained only 30% of Texas observations. This provides an indication of the challenge of differentiating path vs. non-path in FL compared to TX. Moreover, as you’ll see, it doesn’t really matter in this case.⁵

 

3.C. Comparison of Risk-adjusted Default Rates

Consistent with the above time series, we see that path default rates uniformly exceed non-path rates, which uniformly exceed hurricane-free rates, when adjusted for risk.

Notice the nice, increasing, convex function formed by the green dots; that’s why we like our Spero Portfolio Risk Measure. Also notice that if observations in the red and blue dots were grouped with the green dots, the relationship would be weaker, and that’s why we control for them.

3.D. Comparison of Risk-adjusted LGDs

As with Harvey in TX, with Irma in Florida we see that when we control for risk, LGDs in hurricane-free times exceed path and non-path LGDs. In fact, LGDs in Florida in the quarters immediately after Irma are basically zero.

3.E. Comparison of Risk-adjusted Charge-off Rates

Well, regardless of how high default rates are−even if they are 100%−when you multiply them by LGDs of zero, it should be little surprise that overall loss rates are… zero. Here, when risk is controlled, path, non-path, and non-hurricane observations are indistinguishable from each other: almost nothing is lost until risk increases.

Given that overall loss rates are indistinguishable from each other in the relevant domain−about [35, 90] on the horizontal axis−there would be little justification for any type of hurricane-related overlay. However, for transparency and conceptual soundness, which we define as representational faithfulness and reliability, we would control for the differences in default rates and LGDs and forecast each based on the curves formed by the green dots in their respective graphs. Once again, notice the beautiful increasing convex function of charge-offs w.r.t. our risk measure.

Another way to visualize net charge-offs is through the interactive heatmap below, which reinforces the conclusions we came to from our scatterplot above. Note that observation counts can vary from tile to tile, and ranges with less than 10k observations are filtered out.

4. Hurricane Katrina, Louisiana, 2005

4.B. Default Rates in Louisiana through Time

Unlike Harvey and Irma in 2017, Katrina hit a few years before The Financial Crisis, and towards the beginning of our mortgage data history. Below, we see peak default rates for Katrina to be four-to-six times higher than other  two storms.⁶Like Irma, we see the blue non-path rates are a bit higher than the default rates in green, contiguous hurricane-free times; so, perhaps we should have widened the path.⁷

4.C. Comparison of Risk-adjusted Default Rates

We see that default rates for mortgages in the path greatly exceed non-path rates, which are barely higher than hurricane-free rates.

Like Harvey and Irma, we see a fairly strong (increasing) relationship for default rates as a function of risk for mortgages in the path. What’s striking here−compared to Texas or Florida−is that default rates for hurricane-free times don’t spike as dramatically as the risk measure increases. That’s partly a scaling effect, but it is also likely due to the fact that Katrina hit before the crisis, and Louisiana did not see the frothy, bubbly housing market that other Sunbelt states saw.

4.D. Comparison of Risk-adjusted LGDs

As with the other two states, we see that hurricane-free LGDs exceed both path and non-path LGDs, which seem to be similar. Unlike default rates, for hurricane-free observations, LGDs show a sharper increase as our risk measure peaks.

4.E. Comparison of Risk-adjusted Charge-off Rates

For Harvey, we saw loss rates in the path to be about two basis points (2 BPs) higher than the other two categories, while for Irma there was no difference; all were close to zero. For Katrina, however, we see a much larger difference of about six basis points (6 BPs).

So, the question arises: for a future hurricane, should one anticipate higher losses, closer to Katrina, or much smaller differences, like those for Harvey and Irma? (There was absolutely no difference for Irma.)

Another way to visualize net charge-offs is through the interactive heatmap below. Note that observation counts can vary from tile to tile, and ranges with less than 10k observations are filtered out.

We would argue that after the crisis, both underwriting and servicing standards became stricter−both at Fannie and at banks−so those 2017 storms are more representative of possible losses than the (or a) 2005 storm.

Do we have any evidence of that? Well, yes, a bit, and that involves a comparison of Louisiana flooding pre- and post-crisis. (The post-crisis storm was not as severe as Katrina−or it would have an infamous name.) We understand that one could make a convexity argument that losses grow disproportionately to, say, wind speed or water depth, but the flooding did occur in some of the same parishes as Katrina; and, it is our best, readily available comparison.

5. Flooding in Louisiana, 2016

5.B. Default Rates in Louisiana through Time

To create this time series, we eliminated defaults within Katrina’s path but left the blue non-path defaults as part of the sequence of “Normal Times” observations. Again, with a wider Katrina path, those blue spikes on the left would be lower, but they’re not the focus of this section’s analysis; so, no worries.

Here, it seems like we did a good job capturing the flooded area by parish as the blue bars on the right are close to the surrounding green ones. The comparison worth noting here is that the worse quarter for 2016 default rates was almost twice the peak of crisis-related default rates in 2010. We know that’s not true across the crisis; it is only a single quarter, but it is still striking.

5.C. Comparison of Risk-adjusted Default Rates

As with the hurricanes, we see that default rates for mortgages in the path greatly exceed those not in the path, which are about the same as observation in hurricane-and-flood-free times.

Within the flood zone, default rates are non-monotonic with our generic risk measure because of the red point to the left. We still see the anticipated convex relationship between our risk measure and defaults during “normal times.”

From a casual, visual inspection of the scatter plot, it appears that flood zone default rates are about three times the other two categories on a risk-adjusted basis.

5.D. Comparison of Risk-adjusted LGDs

Once again, we see that LGDs during normal weather substantially exceed those in the flood zone−a cursory glance says by about a factor of three when adjusted for risk.

5.E. Comparison of Risk-adjusted Charge-off Rates

Well, flood zone defaults rate are about three times higher than the other two categories and LGDs for the other two categories are about three times those in the flood zone; so, when you do the multiplication, there turns out be no real difference in loss rates, which is what the scatter plot shows.

Again, we provide an interactive heatmap that shows the same conclusions are our scatterplot.  Note that observation counts can vary from tile to tile, and ranges with less than 10k observations are filtered out.

 

6. Hurricane Katrina, Mississippi, 2005

6.B. Default Rates in Mississippi through Time

 Below, we see the same peak default rates for Katrina in Mississippi to be four-to-six times higher than the other two storms ( and slightly lower than Katrina in LA).⁶Like Irma and Katrina in LA, we see the blue non-path rates are a bit higher than the default rates in green, contiguous hurricane-free times; so, perhaps we should have widened the path.⁷

6.C. Comparison of Risk-adjusted Default Rates

We see that default rates for mortgages in the path greatly exceed non-path rates, which are barely higher than hurricane-free rates.

We come to the same conclusions we did for Katrina in Louisiana. Like Harvey and Irma, we see a fairly strong (increasing) relationship for default rates as a function of risk for mortgages in the path. Again, what is striking here compared to Texas or Florida−is that default rates for hurricane-free times don’t spike as dramatically as the risk measure increases. That’s partly a scaling effect, but it is also likely due to the fact that Katrina hit before the crisis, and Mississippi did not see the frothy, bubbly housing market that other Sunbelt states saw.

6.D. Comparison of Risk-adjusted LGDs

Once again, we see that LGDs during normal weather exceed those in the flood zone.

6.E. Comparison of Risk-adjusted Charge-off Rates

In general, we see a similar outcome to the other natural disasters that have been covered so far.

Again, we provide an interactive heatmap that shows the same conclusions are our scatterplot.  Note that observation counts can vary from tile to tile, and ranges with less than 10k observations are filtered out.

We can ask the same question for Katrina in Mississippi that we did in Louisana: for a future hurricane, should one anticipate higher losses, closer to Katrina, or much smaller differences, like those for Harvey and Irma? (There was absolutely no difference for Irma.)
Given (1) differences in underwriting and servicing standards pre-and post-crisis, and (2) evidence from the 2016 flooding, it’s not clear that Katrina’s higher loss rates would reoccur today or tomorrow.

7. Summary

So, what are our main takeaways from each storm?

• For Irma in Florida in 2017 and the widespread flooding in Louisiana in 2016, there was no difference in loss rates for mortgages in the path compared to mortgages outside of the path or observations from hurricane-free times.
• For mortgages on houses in the path of Harvey in Texas in 2017, losses were about two (2) BPs higher than in the rest of Texas or during hurricane-free times.
• For Katrina in 2005, losses averaged about six (6) BPs higher than the rest of Louisiana or during hurricane-and-flood-free times. Given (1) differences in underwriting and servicing standards pre-and post-crisis, and (2) evidence from the 2016 flooding, it’s not clear that Katrina’s higher loss rates would reoccur today or tomorrow.

What can we conclude based on the four storms?

• Mortgages in the storm path have higher default rates than those outside of the path or observations from time periods without severe storms; so, they should be accounted for in default models.
• LGDs during storms are lower than those during “normal”, hurricane-free times by roughly the proportion default rates in storm paths are higher. Thus, they should be accounted for in LGD models.
• Combining the last two bullets, we conclude that when default rates are multiplied by LGDs to get loss rates, the net effect is close to zero.

If you have questions or comments or would like to learn more about our loss forecasting methods and experience, please contact us.

Footnotes