Household financial vulnerabilities and physical climate risks

Introduction

Along with increases in global temperatures, natural disasters happen more frequently than they did in the past.1 For households exposed to severe weather events, this creates financial risks for two possible reasons:

  • depreciation in the value of assets when houses that are not fully insured are destroyed
  • loss of income if employees cannot reach their workplace or businesses are disrupted

As a result of these risks, natural disasters can trigger financial distress among severely affected households. Such distress was observed in Fort McMurray, Alberta, when the share of borrowers who fell behind on their mortgage payments rose sharply after the 2016 wildfires.2

We examine the intersection between severe weather events and household financial vulnerabilities. Certain regions across Canada have historically been more affected by natural disasters. Some of these areas also have a larger share of financially vulnerable households, which can potentially amplify the losses for the wider financial system. Better understanding the overlap between these two concepts is important for evaluating the financial system’s resilience to physical climate risks.

Measuring exposure to natural disasters

We use historical data to identify areas most exposed to natural disasters (Chart 1). The Canadian Disaster Database compiled by Public Safety Canada allows us to track disasters that meet criteria outlined in the Emergency Management Framework for Canada. A disaster is defined as a hazard that affects a vulnerable community ”in a way that exceeds or overwhelms that community’s ability to cope and may cause harm to the safety, health, welfare, property or environment of those people”.3 To be included in the database, a disaster event must meet at least one of the following criteria:

  • 10 or more people killed
  • 100 or more people injured, evacuated or left homeless
  • an appeal for national or international assistance
  • historical significance
  • significant damage or interruption affecting the community’s ability to recover on its own

Chart 1: The frequency of natural disaster events has increased over time

Note: Some of the increase in the number of natural disasters over time likely reflects more systematic reporting and better measurement by authorities.
Sources: Canadian Disaster Database and Bank of Canada calculationsLast observation: 2016

We consider 15 categories of natural disasters, which we map to forward sortation areas (FSAs).4, 5 When considering the total frequency of events across all types of natural disasters, we find that disasters are most frequent in the Prairies. However, simply using the frequency of disasters to sort the most exposed areas could be misleading. For example, an FSA strongly exposed to wildfires would not be classified as strongly exposed to disasters compared with an FSA exposed to floods because wildfires occur less frequently than floods. Also, natural disaster events are not always independent. For example, extreme drought, a disaster itself, can also increase the likelihood of wildfires.

We instead identify areas most exposed to natural disasters by creating a multi-hazard exposure index following Dilley et al. (2005). We proceed in two steps. First, we rank FSAs into low, medium or high exposure for each natural disaster category based on the frequency distribution of events:

  • low exposure—at or below the 40th percentile
  • medium exposure—between the 40th and 80th percentiles
  • high exposure—at or above the 80th percentile

Second, we combine this ranking across all disaster types into a multi-hazard exposure index. In particular, based on their highest exposure across all 15 disaster types, FSAs are classified into the following index categories:

  • low exposure across all disasters
  • medium exposure to at least one type of disaster
  • high exposure to at least one type of disaster—this category is further split into those with high exposure to one, two or more natural disasters

Most regions have high exposure to at least one type of natural disaster (Chart 2). FSAs with high exposure to multiple types of disasters are concentrated in British Columbia, the Prairies, the Atlantic provinces, Northern Quebec, the Northwest Territories and parts of Southern Ontario, including Toronto. Generally, smaller FSAs represent areas with denser populations and a higher concentration of physical assets.

Chart 2: Exposure to different types of natural disasters varies by region across Canada

Chart 2: Exposure to different types of natural disasters varies by region across Canada

Multi-hazard exposure index, by forward sortation area

Exposure to different types of natural disasters varies by region across Canada

Data available as: CSV, JSON and XML
Sources: Canadian Disaster Database and Bank of Canada calculations

Description: Exposure to different types of natural disasters varies by region across Canada

Chart 2 is a map of Canada showing the relative exposure of forward sortation areas to natural disasters.

The map builds on a multi-hazard exposure index that ranges from shades of yellow to orange to dark red. The values are defined as follows:

  • yellow – Low exposure to all disasters
  • light orange – Medium exposure to 1 or more disasters
  • orange – High exposure to 1 disaster
  • light red – High exposure to 2 disasters
  • dark red – High exposure to 3 or more disasters

High-exposure areas (light and dark red) are concentrated in British Columbia, the Prairie provinces, Northern Quebec, the Northwest Territories and parts of Southern Ontario, including Toronto.

Measuring household financial vulnerabilities

We compile two sets of indicators of household financial vulnerability by FSA. For each FSA, we aggregate anonymized loan-level data from TransUnion and mortgage origination data from the regulatory filings of Canadian banks.6

We consider indicators of debt burden, focusing on mortgage debt because housing is typically the largest asset on a household’s balance sheet and is often accompanied by a mortgage loan. The physical destruction of homes can therefore be an important source of loss in the financial system. We include:

  • the share of mortgages for which the loan is at least 4.5 times the household’s income (loan-to-income ratio above 450 percent)
  • the average loan-to-income ratio of mortgages
  • the share of borrowers with a mortgage

We also consider indicators that measure how limited households’ access to credit is. Households with less access to credit could find it more difficult to cope with sudden disruptions to their incomes during a disaster. We include:

  • the share of mortgages with mortgage default insurance7
  • the share of borrowers who use more than 90 percent of their available credit on credit cards or lines of credit
  • the share of borrowers with below-prime credit scores

Using machine learning to find vulnerability patterns

To identify financially vulnerable FSAs exposed to physical risks, we rely on unsupervised machine learning. Rather than identifying these FSAs using arbitrary thresholds, a machine learning algorithm finds the most common patterns of financial vulnerability and multi-hazard exposure among FSAs. We follow two steps:

We first reduce dimensionality by using principal component analysis.8 This allows us to simplify our classification problem from seven indicators to three dimensions, thereby improving the ability of the clustering algorithm to identify distinct groups, or clusters, within the data.

We then identify three clusters of similar FSAs with a k-means clustering algorithm.9 This statistical procedure finds groups in a dataset by minimizing the distance between data points. In our case, we identify three clusters of FSAs with similar characteristics in terms of multi-hazard exposure and financial vulnerabilities. We confirm the existence of three clusters using scree and silhouette coefficient plots (Chart 3).10

Chart 3: The unsupervised machine learning algorithm identifies three clusters of data

Chart 3: The unsupervised machne learning algorithm identifies three clusters of data


a. The squared distance of each point to the centre of the cluster is small

b. The average distance of points in a cluster to the nearest cluster is large

Sources: Canadian Disaster Database, TransUnion, regulatory filings of Canadian banks and Bank of Canada calculations

We analyze the three clusters to provide an economic interpretation (Table 1). The machine learning algorithm can classify data points into clusters based on several dimensions of similarities, but it cannot interpret them. We therefore examine each cluster’s summary statistics and characterize the three clusters as follows:

  • cluster A—higher exposure to disasters and higher indebtedness
  • cluster B—moderate exposure to disasters and limited access to credit
  • cluster C—lower exposure to disasters and lower financial vulnerabilities

Table 1: Characteristics of forward sortation areas in each of the three clusters identified through machine learning

    Clusters
Concepts Individual indicators A B C
Physical climate risk FSAs with high exposure to three or more disasters (%) 68 25 6
FSAs with high exposure to two disasters (%) 5 48 9
FSAs with low to high exposure of no more than one disaster (%) 27 27 85
Debt burden vulnerability Share of loan-to-income ratios above 450% (%) 28 10 11
Average loan-to-income ratio (%) 339 233 259
Borrowers with mortgages (%) 27 28 30
Credit constraint vulnerability Insured mortgages (%) 10 35 22
Use of more than 90% of available credit on credit cards or lines of credit (%) 13 19 14
Borrowers with below prime credit scores (%) 19 26 18

Note: FSA means forward sortation area. The values in this table represent shares or averages across the grouped FSAs in 2020Q3. Clusters with the highest vulnerability based on each indicator are colour-coded red. Those with the second-highest vulnerability are colour-coded yellow. Those with the lowest vulnerability are colour-coded green.

Assessing the overlap between disaster exposure and financial vulnerabilities

Households in FSAs with high indebtedness and high exposure to natural disasters hold 39 percent of Canadian household debt and are mainly located in British Columbia and Ontario (Chart 4, red bar, corresponds to cluster A). Large debt burdens are mostly linked to elevated house prices. In this case, the destruction of leveraged physical assets such as housing could amplify financial system losses.

Chart 4: Households that are both financially vulnerable and exposed to natural disasters hold a large fraction of household debt

Sources: Canadian Disaster Database, TransUnion and Bank of Canada calculationsLast observation: 2020Q3

Households in FSAs with moderate exposure to natural disasters and limited access to credit hold 17 percent of Canadian household debt and are mostly located in the Prairies and Atlantic provinces (Chart 4, yellow bar, corresponds to cluster B). Households located in these FSAs may be less able to cope with shocks by using their lines of credit because natural disasters can lead to a sudden disruption in income due to, for example, displacement or the destruction of means of production.

Lastly, households that hold the remaining 44 percent of Canadian household debt are in FSAs primarily located in Quebec and, to a lesser extent, in Ontario and the Prairies (Chart 4, green bar, corresponds to cluster C). On average, households in these FSAs have relatively low exposure to natural disasters, and their financial vulnerabilities may be less likely to amplify the consequences of natural disasters.

Conclusion and limitations

Our analysis highlights regions where the combination of household financial vulnerabilities and exposure to natural disasters could amplify financial losses. This suggests that individual households’ characteristics are important factors for climate stress tests to assess whether an exposure to physical climate risk is also a vulnerability for the financial system. Still, three main limitations of our work should be kept in mind:

  • We rely on historical data, which does not provide information about how physical risks will evolve. The frequency, severity and regional distribution of disasters could change over time for various reasons, ranging from climate change to population growth.11
  • We lack comprehensive data on households’ balance sheets. For households with ample liquid assets, exposure to natural disasters may imply a lower financial risk than suggested in our analysis.
  • We lack information on households’ insurance coverage. However, even if households have insurance, standard home insurance might not provide coverage for all types of disasters.

Endnotes

  1. 1. See Amano, Gosselin and Mc Donald-Guimond (2021) and IPCC (2013; 2007).[]
  2. 2. See Bilyk et al. (2020) and Ho et al. (forthcoming).[]
  3. 3. See Public Safety Canada’s Canadian Disaster Database.[]
  4. 4. Categories include avalanches, cold events, droughts, earthquakes, floods, heat events, hurricanes, landslides, storm surges, severe thunderstorms, tornadoes, tsunamis, wildfires, winter storms and unspecified storms. We exclude certain categories of disasters included in the Canadian Disaster Database, such as geomagnetic storms, epidemics and infestations, because the associated events do not constitute a physical risk of climate change that can affect households through the destruction of physical assets.[]
  5. 5. Forward sortation areas (FSAs) are geographical areas identified by the first three characters of their postal code. We include 1,658 FSAs.[]
  6. 6. To protect the privacy of Canadians, TransUnion did not provide any personal information to the Bank. The TransUnion dataset was anonymized, meaning that it does not include information that identifies individual Canadians, such as names, social insurance numbers or addresses. In addition, the dataset has a panel structure, which uses fictitious account and consumer numbers assigned by TransUnion.[]
  7. 7. Borrowers who have a mortgage with a loan-to-value ratio greater than 80 percent are required to purchase mortgage default insurance. This insurance protects lenders if a borrower defaults on the mortgage loan.[]
  8. 8. See Ding and He (2004).[]
  9. 9. See Kaufman and Rousseeuw (1990).[]
  10. 10. See Rousseeuw (1987).[]
  11. 11. See Ens and Johnston (2020) for a forward-looking assessment of the potential macroeconomic impact of climate change in Canada.[]

References

  1. Amano, R., M.-A. Gosselin and J. Mc Donald-Guimond. 2021. “Evolving Temperature Dynamics in Canada: Preliminary Evidence Based on 60 Years of Data.” Bank of Canada Staff Working Paper No. 2021-22.
  2. Bilyk, O., A. T. Y. Ho, M. Khan and G. Vallée. 2020. “Household Indebtedness Risks in the Wake of COVID‑19.” Bank of Canada Staff Analytical Note No. 2020-8.
  3. Dilley, M., R. S. Chen, U. Deichmann, A. Lerner-Lam, M. Arnold, J. Agwe, P. Buys, O. Kjevstad, B. Lyon and G. Yetman. 2005. Natural Disaster Hotspots: A Global Risk Analysis. Washington, D.C.: World Bank Group.
  4. Ding C. and X. He. 2004. “Cluster Structure of K-means Clustering via Principal Component Analysis.” Advances in Knowledge Discovery and Data Mining Lecture Notes in Computer Science 3056: 414–418.
  5. Ens, E. and C. Johnston. 2020. “Scenario Analysis and the Economic and Financial Risks from Climate Change.” Bank of Canada Staff Discussion Paper No. 2020-3 (May 2020).
  6. Ho, A. T. Y., K. P. Huynh, D. Jacho-Chavez and G. Vallée. Forthcoming. “We Didn’t Start the Fire! The Effects of a Natural Disaster on Consumer Financial Distress.” Bank of Canada Staff Working Paper.
  7. Intergovernmental Panel on Climate Change (IPCC). 2007. “Summary for Policymakers.” In Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, edited by S. Solomon, D. Qin, M. Manning, M. Marquis, K. Averyt, M. B. Tignor H. L. Miller and Z. Chen, 1–18. Cambridge, United Kingdom, and New York, New York: Cambridge University Press.
  8. Intergovernmental Panel on Climate Change (IPCC). 2013. “Summary for Policymakers.” In Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, edited by T. F. Stocker, D. Qin, G.-K. Plattner, M. Tignor, S. K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex and P. M. Midgley, 3–29. Cambridge, United Kingdom, and New York, New York: Cambridge University Press.
  9. Kaufmann, L. and P. J. Rousseeuw. 1990. Finding Groups in Data: An Introduction to Cluster Analysis. New York, NY: John Wiley & Sons.
  10. Rousseeuw, P. J. 1987. “Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis.” Journal of Computational and Applied Mathematics 20 (November): 53–65.

Disclaimer

Bank of Canada staff analytical notes are short articles that focus on topical issues relevant to the current economic and financial context, produced independently from the Bank’s Governing Council. This work may support or challenge prevailing policy orthodoxy. Therefore, the views expressed in this note are solely those of the authors and may differ from official Bank of Canada views. No responsibility for them should be attributed to the Bank.