When Natural Disasters Test Underwriting Resilience: Evidence from Vietnam’s Non-Life Insurance Sector

Natural disasters pose a dual challenge for insurance markets. They can increase demand for protection by making climate-related losses more visible, but they can also expose weak underwriting, pricing and risk management. This matters as extreme weather becomes more frequent and severe, and as insurers are expected to support recovery from climate-related losses [1, 2].

Prior research shows that physical climate risk affects economic activity, firm performance and financial decisions [3-5].

For insurers, the channel is direct: they underwrite the risks affected by disasters. Insurance prices, pools, and transfers risk and can reduce post-disaster financial stress [6]. But severe or uncertain disasters may raise claims, weaken pricing accuracy, pressure reserves and increase reinsurance needs. Evidence from Chinese property-casualty insurers shows that climate risks can increase claim ratios and reduce performance when risk management is insufficient [7].

This makes insurance relevant to climate-adaptation planning and sustainable development. In climate-vulnerable economies, a resilient non-life insurance sector helps households, firms and governments absorb losses, reduce post-disaster fiscal pressure and maintain economic continuity.

Disaster losses can therefore be read as demand shocks and as stress tests of underwriting quality. Prior studies show that disaster experience can increase insurance take-up [8-10]. Yet higher demand does not guarantee sustainable premium growth if insurers underprice risk, hold loss-prone portfolios or control claims poorly.

Vietnam is a useful setting for this question. It is exposed to storms, floods and heavy rainfall, while its non-life insurance market is still developing and faces low penetration and protection gaps [11-13]. Under the Law on Insurance Business No. 08/2022/QH15 [14], solvency management, risk governance and prudential supervision have become more salient. This setting links insurer resilience to disaster-risk financing, climate adaptation and the objectives of SDG 13 and SDG 11.

Vietnam’s vulnerability is especially important because climate risk is transmitted through sectors that are central to the country’s development model, including agriculture, transport, trade, industry, energy, coastal infrastructure and the Mekong Delta. The World Bank’s Country Climate and Development Report highlights that Vietnam has more than 3,200 km of coastline, many low-lying cities and river-delta regions, and is among the countries most vulnerable to climate change. Initial estimates suggest that climate-change impacts caused losses of about US$10 billion in 2020, equivalent to 3.2% of GDP. Without adequate adaptation and mitigation measures, climate change could cost Vietnam about 12–14.5% of GDP annually by 2050 [11]. These figures show that climate risk in Vietnam is not only a physical threat to production, infrastructure and livelihoods, but also a financial risk that can be transmitted through insurance demand, claims payments, reserve adequacy, reinsurance needs and insurers’ underwriting capacity. This makes Vietnam a particularly relevant setting for examining whether natural disaster-loss conditions reveal underwriting vulnerability in the non-life insurance sector.

Existing studies mainly estimate average climate-risk effects on insurers or firms. Ding and Li [7] showed that climate risks increase claim ratios and weaken performance among Chinese property-casualty insurers. Less is known about whether disaster-loss conditions reveal pre-existing underwriting vulnerability across insurers, especially in emerging markets with pricing constraints, data limitations and protection gaps.

This study examines whether aggregate disaster-loss conditions reveal heterogeneous underwriting vulnerability among Vietnam’s listed non-life insurers. Using firm-level panel data for 2014-2024, it interacts national disaster losses with insurers’ prior claims ratios and studies five outcomes: premium revenue, claims ratio, return on equity (ROE), reserve ratio and reinsurance ratio. Because disaster losses vary only by year, the design does not estimate insurer-specific catastrophe exposure or the average causal effect of disasters. It asks whether insurers with different prior claims performance respond differently under the same national disaster-loss environment.

Two limitations guide the interpretation. The sample covers listed non-life insurers, which are dominant but not the whole market. Disaster losses are national-year measures, so the results are interpreted as heterogeneous associations under common disaster-loss conditions.

The main result is that insurers with higher prior claims ratios show weaker premium-side adjustment when aggregate disaster losses increase. Evidence of stronger claims pressure is supportive but less robust, while ROE, reserves and reinsurance show no clear differential response. The findings suggest that disaster-loss conditions reveal underwriting vulnerability mainly through premium-side adjustment.

The study makes three contributions. It shifts the analysis from average climate-risk effects to heterogeneous underwriting vulnerability; provides evidence from Vietnam’s climate-exposed and developing non-life insurance market; and shows that prior claims performance can serve as an observable early-warning indicator. These contributions connect insurer resilience with sustainable disaster-risk financing and climate-adaptation planning. The rest of the paper is organized as follows. Section 2 reviews the literature and develops the hypotheses. Section 3 describes the data and methodology. Section 4 presents the results and discussion. Section 5 reports robustness and small-sample sensitivity tests. Section 6 discusses contributions and policy implications. Section 7 concludes.

3.1 Data and sample

This study uses firm-level panel data from publicly listed non-life insurance firms in Vietnam. The initial dataset covers listed insurers over the period 2008–2024. BVH is excluded from the main analysis because its consolidated financial statements combine life insurance, non-life insurance and other financial activities, making it less comparable to firms operating primarily in the non-life insurance segment.

The focus on non-life insurers is intentional. Unlike firms in the broad corporate market, non-life insurers are more directly exposed to disaster-related underwriting risks through their core insurance activities. Natural disasters may increase insurance demand by raising risk awareness, but they may also generate higher claims, expose underpricing, test reserve adequacy and affect the need for reinsurance. Therefore, the non-life insurance sector provides an economically meaningful setting in which disaster-loss conditions can be examined through observable underwriting outcomes such as premium revenue, claims ratios, technical reserves and reinsurance. A broad-market sample would mix firms with different and often indirect channels of climate-risk exposure, potentially diluting the underwriting mechanism that this study seeks to identify.

The main empirical analysis focuses on the period 2014–2024, during which all 10 listed non-life insurers are consistently represented in the dataset. This results in a balanced panel of 110 firm-year observations. The 2014–2024 period is selected as the baseline sample because it provides a stable firm composition and reduces concerns that the results are driven by changes in sample coverage in earlier years. Because the regression models use one-year lagged firm-level variables, the first year of the baseline panel is lost in the main estimations. Additional missing observations in specific accounting items further reduce the effective sample to 99–100 observations depending on the specification.

Although the number of firms is relatively small, the sample is economically meaningful because it captures a substantial share of Vietnam’s non-life insurance market. From 2014 to 2024, the sample firms collectively account for approximately 56–70% of total non-life insurance premiums, with an average market coverage of 62.69%. This reflects the concentrated structure of Vietnam’s listed non-life insurance sector and supports the interpretation of the sample as evidence from dominant listed market participants. Nevertheless, the results should be interpreted as evidence on listed non-life insurers’ resilience rather than as representative of the entire non-life insurance market, especially smaller unlisted insurers.

To assess whether the findings are sensitive to the baseline period, the full unbalanced panel from 2008 to 2024 is used in robustness checks. After excluding BVH, this broader sample includes 139 firm-year observations. However, because the earlier years contain fewer consistently observed firms, the balanced 2014–2024 panel is used as the preferred sample for the main analysis.

Financial statement data are obtained from FiinGroup’s FiinPro standardized financial statement database for listed Vietnamese firms [27]. FiinPro reports standardized annual financial statement items in VND, scaled to billion VND in the exported tables. The extracted items are mapped to the corresponding accounting items in insurers’ financial statements, including premium revenue, claims expense, profit after tax, equity, total assets, technical reserves and reinsurance expenses. The exact Vietnamese financial statement items, English translations, source statements, formulas, units and data-treatment rules used to construct each variable are reported in Appendix B.

Data on monetary losses caused by natural disasters are collected from the official annual disaster-damage statistical tables published in the ‘Thống kê thiệt hại’ section of Vietnam’s disaster management authority [28]. For each year, the study uses the annual aggregate monetary loss reported in the final monetary-loss row of the corresponding annual table. The original figures are reported in million VND and are converted into billion VND by dividing by 1,000 before applying the natural logarithm. All disaster-loss values are nominal and are not adjusted for inflation. Therefore, the disaster-loss variable should be interpreted as reported aggregate nominal disaster-loss conditions rather than real disaster-loss exposure or insurer-specific catastrophe exposure.

Table 1 reports the annual market coverage of the baseline sample.

Table 1. Market coverage of the baseline sample

Note: Market share is calculated as the sum of sample firms’ premium revenue divided by total non-life insurance premium revenue in Vietnam.

3.2 Variable definitions and rationale

Table 2 summarizes the variable definitions, measurements and empirical roles.

Table 2. Variable summary

The variable definitions follow prior studies on climate risk and insurer performance, while being adapted to the Vietnamese data context. Premium revenue, claims ratios and profitability are used to capture insurers’ underwriting scale, claims pressure and overall financial performance. Ding and Li [7], for example, used premium income, claims expenditure, claim ratios, ROA and ROE to examine how climate risks affect property-casualty insurers. Claims ratios are particularly important in this study because prior evidence suggests that claims are a primary channel through which climate risk affects insurer performance.

Reserve and reinsurance ratios are included to examine whether insurers adjust internal risk-absorption or external risk-transfer arrangements when disaster losses increase. These variables are treated as exploratory risk-management response outcomes rather than confirmed mechanisms. This treatment is consistent with prior evidence showing that insurers do not always significantly adjust reinsurance or business distribution in response to climate risk.

The disaster-loss variable is constructed from official national disaster-damage statistics and captures aggregate disaster severity in Vietnam. Unlike studies that measure climate exposure at the city-insurer or firm-location level, this variable varies only at the national-year level. Therefore, it should be interpreted as a measure of aggregate disaster-loss conditions rather than insurer-specific catastrophe exposure. Building on the evidence that claims are a central channel of insurer vulnerability, the lagged claims ratio is used as an observable proxy for prior underwriting vulnerability. The interaction between aggregate disaster losses and lagged claims ratios is the main coefficient of interest and captures whether insurers with weaker prior claims performance exhibit different responses when disaster losses increase.

3.3 Empirical strategy

The empirical strategy is designed to examine whether insurers with different prior underwriting conditions exhibit different financial responses when aggregate natural disaster losses increase. Because the disaster-loss variable is measured at the national-year level, it does not vary across firms within the same year. In specifications with year fixed effects, the standalone effect of Log_damage is therefore absorbed by year effects. Accordingly, the analysis focuses on interaction specifications that combine aggregate national disaster losses with lagged firm-level exposure variables.

The main specification is written as follows:

${{Y}_{it}}={{\beta }_{0}}+{{\beta }_{1}}\left( \text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}\times \text{lag }\!\!\_\!\!\text{ claims }\!\!\_\!\!\text{ rati}{{\text{o}}_{i,t-1}} \right)+{{\beta }_{2}}\text{ }\!\!~\!\!\text{ lag }\!\!\_\!\!\text{ claims }\!\!\_\!\!\text{ rati}{{\text{o}}_{i,t-1}}+{{\beta }_{3}}{{X}_{i,t-1}}+{{\mu }_{i}}+{{\lambda }_{t}}+{{\varepsilon }_{it}}$

where, ${{Y}_{it}}$ represents the financial outcome of insurer i in year t. The model is estimated separately for five dependent variables: log_premium, claims_ratio, ROE, reserve_ratio and reinsurance_ratio. These outcomes capture premium-side adjustment, underwriting pressure, overall profitability, internal risk absorption and external risk transfer, respectively.

$\text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}$ is the logarithm of aggregate nominal monetary losses caused by natural disasters in Vietnam in year t.$\text{ }\!\!~\!\!\text{ lag }\!\!\_\!\!\text{ claims }\!\!\_\!\!\text{ rati}{{\text{o}}_{i,t-1}}$ is the previous-year claims ratio of insurer i and is used as an observable proxy for prior underwriting vulnerability. The interaction term $\text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}\times \text{ }\!\!~\!\!\text{ lag }\!\!\_\!\!\text{ claims }\!\!\_\!\!\text{ rati}{{\text{o}}_{i,t-1}}$ is the main variable of interest. Its coefficient, ${{\beta }_{1}}$, captures whether insurers with higher prior claims ratios exhibit different financial responses to the same aggregate disaster-loss conditions than insurers with lower prior claims ratios.

${{X}_{i,t-1}}$ is a vector of lagged firm-level controls, including firm size, lagged reserve ratio and lagged reinsurance ratio. Firm fixed effects, ${{\mu }_{i}}$, control for time-invariant differences across insurers, such as ownership structure, business model, underwriting capacity and distribution network. Year fixed effects, ${{\lambda }_{t}}$, control for common shocks affecting all insurers in the same year, including macroeconomic conditions, regulatory changes, inflation, insurance market growth and other nationwide factors.

Because $\text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}$ varies only by year, its standalone effect is absorbed by year fixed effects. The coefficient of interest is therefore not the direct coefficient of $\text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}$, but the interaction coefficient ${{\beta }_{1}}$. The estimates should be interpreted as evidence of heterogeneous firm-level responses to aggregate disaster-loss conditions, rather than as direct estimates of insurer-specific catastrophe exposure or average disaster effects. This identification choice is important: the study does not claim to measure each insurer’s geographic or product-line catastrophe exposure, but tests whether insurers with different prior underwriting conditions respond differently under the same national disaster-loss environment.

The expected interpretation of ${{\beta }_{1}}$ depends on the outcome variable. In the premium model, a negative coefficient would suggest that insurers with higher prior claims ratios exhibit weaker premium-side adjustment when disaster losses increase. In the claims-ratio model, a positive coefficient would indicate stronger underwriting pressure among these insurers. In the ROE model, the sign is treated as an empirical question because profitability reflects both underwriting and non-underwriting activities. In the reserve-ratio and reinsurance-ratio models, significant coefficients would suggest that insurers with higher prior claims ratios adjust internal buffers or external risk-transfer arrangements in response to disaster-loss conditions. These reserve and reinsurance results are interpreted as exploratory evidence of risk-management responses rather than as confirmed mechanisms.

To examine alternative exposure channels, the study also estimates additional interaction models by replacing $\text{lag }\!\!\_\!\!\text{ claims }\!\!\_\!\!\text{ rati}{{\text{o}}_{i,t-1}}$ with other lagged firm-level characteristics:

${{Y}_{i,t}}={{\beta }_{0}}+{{\beta }_{1}}\left( \text{Log }\!\!\_\!\!\text{ damag}{{\text{e}}_{t}}\times {{Z}_{i,t-1}} \right)+{{\beta }_{2}}{{Z}_{i,t-1}}+{{\beta }_{3}}{{X}_{i,t-1}}+{{\mu }_{i}}+{{\lambda }_{t}}+{{\varepsilon }_{i,t}}$

where, ${{Z}_{i,t-1}}$ represents alternative exposure variables, including lag_market_share, lag_reinsurance_ratio and lag_reserve_ratio. These models examine whether disaster-loss conditions are associated with different responses across insurers with different market positions, prior risk-transfer arrangements or internal buffer capacity.

All firm-level exposure and control variables are lagged by one year to reduce simultaneity concerns between current disaster-loss conditions and firm-level financial responses. Standard errors are clustered at the firm level to account for heteroskedasticity and within-firm serial correlation. However, because the baseline sample contains only 10 firm clusters, conventional firm-clustered standard errors may overstate statistical precision. To address this concern, the baseline results are also evaluated using wild cluster bootstrap p-values for the main interaction term with firm-level clustering. Statistical significance is therefore interpreted with reference to both conventional clustered inference and wild cluster bootstrap inference.

Table 3 summarizes the empirical model structure and exposure variables.

The main analysis is conducted using the balanced panel from 2014 to 2024. The full unbalanced panel from 2008 to 2024 and a sample excluding the COVID-19 years are used as robustness checks. These additional tests examine whether the main findings are sensitive to sample-period selection or to the potential distortion caused by the pandemic years.

Table 3. Empirical model structure and exposure variables

5.1 Robustness checks

Table 7 reports robustness checks using two alternative sample specifications: an extended unbalanced panel covering 2008–2024 and a restricted sample excluding the COVID-19 years. The reported coefficients refer to the interaction term between aggregate disaster losses and prior claims ratios. The coefficient on Log_damage × lag_claims_ratio in the log_premium model remains negative and statistically significant in both specifications, with coefficients of -0.2648 and -0.2762, respectively. This finding supports the main interpretation that insurers with higher prior claims ratios exhibit weaker premium-side adjustment when aggregate disaster losses increase.

The evidence for the claims_ratio model is less stable. The interaction term remains positive and statistically significant in the extended sample, but becomes statistically insignificant after excluding the COVID-19 years, although the sign remains positive. This suggests that the claims-pressure channel is consistent with the baseline interpretation, but should be viewed as less robust than the premium-side result.

The interaction coefficients for ROE, reserve_ratio, and reinsurance_ratio are statistically insignificant across both robustness checks. Overall, Table 7 indicates that the most stable empirical finding is the weaker premium-side adjustment among insurers with higher prior claims ratios. By contrast, the evidence for stronger claims pressure is partial, while the results do not provide strong support for differential profitability, reserve, or reinsurance responses.

Table 7. Robustness check model

Notes: This table reports robustness checks using alternative sample specifications. The reported coefficient is the coefficient on the interaction term Log_damage × lag_claims_ratio. The full-sample specification uses the unbalanced panel from 2008 to 2024, while the excluding-COVID specification removes the COVID-19 years from the baseline sample. Standard errors are clustered at the firm level, and t-statistics are reported in the final column. Statistical significance is based on conventional firm-clustered inference. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively. Wild cluster bootstrap inference is reported for the baseline model in Table 6; Table 7 is interpreted as an additional robustness check.

5.2 Small-sample sensitivity tests

Given the relatively small number of listed non-life insurers in the sample, additional sensitivity tests are conducted to assess whether the main results are driven by model specification or by individual influential firms. Two sets of tests are performed. First, the baseline model is re-estimated using more parsimonious specifications with fewer control variables. Second, leave-one-firm-out tests are conducted by excluding one insurer at a time and re-estimating the full baseline model.

Table A1 reports the results from the parsimonious specifications. Under conventional clustered inference, the claims-ratio result remains positive and statistically significant across all specifications, suggesting that insurers with higher prior claims ratios tend to face stronger underwriting pressure when disaster losses increase. The premium result is more sensitive to the inclusion of firm-level controls. In the minimal specification, the coefficient on Log_damage × lag_claims_ratio is positive but statistically insignificant. However, once firm size is controlled for, the coefficient becomes negative and statistically significant, and remains negative and significant in the full specification. This suggests that firm scale is an important factor in explaining premium-side adjustment and should be controlled for when assessing heterogeneous responses to aggregate disaster-loss conditions across insurers.

Table A2 presents the leave-one-firm-out sensitivity tests. The coefficient on Log_damage × lag_claims_ratio in the premium model remains negative in all ten re-estimations and is statistically significant at the 10% level or better in eight cases. This suggests that the premium-side result is not driven by a single insurer, although its statistical strength is weaker when some firms are excluded. The coefficient in the claims-ratio model remains positive in all ten re-estimations and is statistically significant at the 10% level or better in nine cases. This supports the interpretation that the claims-pressure channel is directionally stable, even though its statistical significance is somewhat sensitive to sample composition.

Overall, these additional sensitivity tests support the main interpretation that aggregate disaster-loss conditions reveal underwriting vulnerability among non-life insurers. The premium-side result is directionally stable across all leave-one-firm-out tests and becomes statistically significant once firm size is controlled for. The claims-pressure result is also directionally stable and remains statistically significant in most leave-one-firm-out specifications. At the same time, the sensitivity of some coefficients reinforces the need to interpret the findings as small-sample evidence from a concentrated emerging insurance market rather than as large-sample causal estimates.

This study examines whether aggregate natural disaster-loss conditions act as a stress test of underwriting resilience among Vietnam’s listed non-life insurers. Using firm-level panel data from 2014 to 2024, the analysis investigates how aggregate economic losses from natural disasters interact with insurers’ prior claims exposure. The study focuses on five financial outcomes: premium revenue, claims ratio, return on equity, reserve ratio, and reinsurance ratio.

The empirical results show that insurers with higher prior claims ratios exhibit weaker premium-side adjustment when aggregate disaster losses increase. This is the most stable finding in the paper and remains marginally significant under wild cluster bootstrap inference, which accounts for the small number of firm clusters. The baseline estimates also provide supportive evidence that these insurers face stronger claims pressure, although this result is weaker under small-cluster inference and less robust across sample specifications. By contrast, the study finds limited evidence that disaster-loss conditions are associated with differential changes in profitability, reserve accumulation, or reinsurance adjustment among more vulnerable insurers. These findings suggest that natural disasters reveal underwriting vulnerability mainly through premium-side adjustment rather than simply creating premium growth opportunities.

The study contributes to research on climate risk, underwriting resilience and sustainable disaster risk financing by showing that prior claims performance matters for insurer responses to aggregate disaster-loss conditions. Claims ratios can serve as practical early-warning indicators, and insurance-sector resilience should be assessed together with claims performance, pricing discipline, reserve capacity and reinsurance arrangements.

From a policy perspective, regulators and insurers should strengthen climate-risk management through underwriting supervision, risk-based pricing, claims monitoring, reserve adequacy assessment and reinsurance stress testing. In Vietnam, these actions can support disaster recovery, climate adaptation and sustainable risk financing, including SDG 11 and SDG 13.

This study has several limitations. The sample is limited to listed non-life insurers and remains relatively small, although these firms represent a substantial share of Vietnam’s non-life insurance market. While wild cluster bootstrap inference is used to address small-cluster concerns in the baseline model, the results should still be interpreted as evidence from dominant listed insurers rather than as representative of the entire insurance market. In addition, natural disaster losses are measured at the national-year level, which does not capture differences in firms’ geographic exposure, product composition, or regional underwriting portfolios. The empirical results should therefore be interpreted as heterogeneous associations under aggregate disaster-loss conditions rather than as insurer-specific catastrophe exposure estimates. Future research could extend this analysis using province-level disaster data, product-line insurance data, longer time-series observations, or firm-level catastrophe exposure measures. Such extensions would allow a more detailed assessment of how climate-related disaster-loss conditions are transmitted through insurance markets and how insurers can strengthen resilience under climate change.