Canadian Disaster Database (CDD)

• The Canadian Disaster Database is a comprehensive disaster-event archive that collects over 1,000 significant incidents since 1900 using defined inclusion criteria.
• It classifies disasters into 13 types and employs Poisson and Generalized Pareto models to quantify frequency and severity for national loss assessment.
• Empirical findings reveal stable disaster frequencies but increasing severity for events like thunderstorm and wildfire, leading to heavier-tailed loss distributions.

Searching arXiv for the specified papers and closely related work on the Canadian Disaster Database. The Canadian Disaster Database (CDD) is a disaster-event archive maintained by Public Safety Canada and used as an empirical basis for analyzing Canadian natural-disaster risk. It contains detailed records on over 1,000 natural, technological, and conflict events (excluding war) since 1900, including events that occurred domestically or abroad but directly affected Canadians. In quantitative research, the CDD has been used both for historical trend analysis and for probabilistic modeling of annual loss distributions, especially for natural disasters related to meteorological or hydrological phenomena (Hao, 5 Oct 2025).

1. Scope, inclusion criteria, and data structure

The CDD records “significant disaster events” rather than all adverse events. Events are included if they meet one or more criteria such as 10 or more fatalities, 100 or more people affected/injured/evacuated/homeless, an appeal for national or international assistance, historical significance, or major disruption such that the affected community cannot recover on its own. The database includes location, date, injuries, evacuations, fatalities, and approximate financial losses. Its data are sourced from traceable references and are reviewed semi-annually (Hao, 5 Oct 2025).

These inclusion rules matter analytically. They imply that the CDD is optimized for consequential events and for national-scale situational awareness, not for a complete census of low-impact local incidents. A plausible implication is that frequency estimates derived from the CDD are estimates for the database’s “significant event” process, rather than for all physically occurring hazards.

2. Event taxonomy and temporal coverage

The CDD classifies natural disasters into 13 types: Avalanche, Cold Event, Drought, Flood, Geomagnetic Storm, Heat Event, Hurricane / Typhoon / Tropical Storm, Storm, Unspecified / Other, Storm Surge, Storms and Severe Thunderstorms, Tornado, Wildfire, and Winter Storm. In the quantitative study focused on natural hazards, the major loss-producing categories are Flood, Thunderstorm, Wildfire, Winter Storm, Tornado, and Storm Other, with smaller categories such as Avalanche, Cold Event, Drought, Hurricane/Typhoon/Tropical Storm, and Storm Surge entering aggregation through average annual losses (Hao, 5 Oct 2025).

The study compares two collections derived from the database: a 2017 collection covering 1900 to 2016 and a 2024 collection covering 1900 to 2020. For stochastic modeling of annual loss frequency and severity, the analysis begins in 1955, because 1955 is the first year with a non-zero loss event recorded in the CDD. The reported totals are 789 events and $22B** in losses for the 2017 collection, and **899 events** and **$34B in losses for the 2024 collection.

The six most frequent disaster categories in the 2024 collection are Flood, Thunderstorm, Wildfire, Winter Storm, Tornado, and Drought; together they account for 86% of all recorded disasters. The top six categories by total loss are Flood, Winter Storm, Thunderstorm, Wildfire, Storm Other, and Drought; together they account for 94% of total disaster-related losses.

Category	Frequency since 1900	Total loss since 1900
Flood	336 events, 37%	$10.0B, 31%
Thunderstorm	141 events, 16%	$5.9B, 17%
Wildfire	113 events, 13%	$4.9B, 15%
Winter Storm	89 events, 10%	$6.7B, 20%
Drought	46 events, 5%	$1.7B, 5%

Within this profile, Flood dominates both by frequency and by cumulative loss, while Winter Storm is notable for high total cost despite lower frequency than flood.

3. Statistical conversion of the CDD into a national loss model

The CDD has been used not only as a descriptive archive but also as input to a compound-loss framework. Event counts are modeled by a Poisson distribution,

$$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$

with

$$\lambda = \frac{\text{number of events in sample period}}{\text{number of years in that period}}.$$

The study gives the example of floods in the 2017 collection: 150 events from 1955–2016, yielding an estimated frequency of $\lambda = 2.46$ (Hao, 5 Oct 2025).

Event severity is modeled with the Generalized Pareto Distribution (GPD),

$$G_{\xi,\beta}(y) = 1 - \left(1 + \xi \frac{y}{\beta}\right)^{-1/\xi},$$

equivalently written in the appendix as

$$F^{GPD}_{\xi,\beta}(x) = 1 - \left(1 + \xi \frac{x}{\beta}\right)^{-1/\xi},$$

where the exceedance is over the threshold of zero-dollar losses, and where $\xi$ is the shape parameter and $\beta$ is the scale parameter.

Annual total loss is then written as

$S = \sum_{k=1}^{K} X_k,$

with $K$ the random number of events in a year and $$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$0 the individual event losses. The characteristic function is expressed as

$$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$1

and, for the Poisson case,

$$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$2

The annual loss distribution is computed by an FFT (Fast Fourier Transform) procedure that discretizes severity, transforms to the frequency domain, applies the compound-loss formula, and then recovers the cumulative distribution by inverse FFT and cumulative summation.

For the aggregate national distribution across event types, the study does not assume perfect independence. Instead, it reports positive Spearman correlations among the six major event types, with the highest around 60%, and uses a Normal copula to model dependence. The procedure simulates marginal annual losses for each event type, maps them through the copula, sums the corresponding annual losses, and repeats this 10,000 times to estimate the aggregate national loss distribution.

4. Empirical findings on frequency, severity, and tail risk

The principal empirical conclusion is that there is no strong evidence that any specific disaster type became more frequent nationwide over the past seven years. This is supported by comparisons of event counts in the 2017 and 2024 collections, by Poisson frequency estimates, and by Wald tests, which are reported as not significant (Hao, 5 Oct 2025).

By contrast, the study identifies a severity shift. The clearest increases are for Thunderstorm and Wildfire, which have caused greater financial losses in recent years, while other categories are described as relatively stable. The parameter estimates reported for the GPD also support this emphasis, especially through larger scale parameters and heavier large-loss behavior in the 2024 collection for these categories.

The national aggregate distribution is interpreted as having become more heavy-tailed. The 2024 collection shows a lower main body of losses at the 25th, 50th, and 75th percentiles, but much higher tail losses at the 90th, 99th, and 99.9th percentiles. This means that ordinary years may look somewhat less costly, while the risk of catastrophic national losses is materially higher.

Percentile ($MM)	2017	2024
25%	108	78
50%	196	164
75%	478	432
90%	1,012	1,173
99%	4,312	5,711
99.9%	6,038	10,460

The study explicitly interprets this pattern as evidence that Canada may be becoming warmer and wetter. The stated reasoning is that Wildfire losses are increasing, consistent with warming and more intense fire conditions; storm-related losses are rising, consistent with wetter and stormier conditions; Storm Other becoming more frequent supports a wetter climate signal; and Drought has declined in relative importance, further supporting a shift away from dryness.

An additional modeling implication concerns low-frequency categories. For some event types, the annual loss distribution begins with a high probability at zero loss, because such events occur only once every few years. The study gives the example that if $$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$3, the probability of zero losses in a year is about 70%.

5. Relation to disaster monitoring and recovery analysis

The CDD is fundamentally an event-cataloguing resource, but its analytical relevance extends to post-disaster measurement and recovery modeling. A separate study using call detail records (CDRs) on the 2010 earthquake in Haiti, the 2015 Gorkha earthquake in Nepal, and Hurricane Matthew in Haiti in 2016 frames disaster analysis in terms of resettlement curves, decay curves, and recovery rates, and explicitly connects such work to the broader problem of documenting displacement and recovery through disaster monitoring and post-disaster analysis (Li et al., 2019).

That CDR-based study characterizes the fraction of internally displaced persons who remain disrupted each week after a disaster and fits the resulting decay with a double-exponential model,

$$f(k) = \frac{e^{-\lambda}\lambda^k}{k!}, \quad k=0,1,2,\dots$$4

interpreted as a mixture of fast-recovering and slow-recovering groups. It reports that half the original number of displaced persons had resettled within four to five weeks across the three disasters, and argues that such methods can assist disaster monitoring, measurement of resilience / recovery, comparison of recovery rates across disasters, and estimation of how many people still require support at a given time.

This connection is significant for understanding the place of the CDD in disaster research. The CDD records event timing, location, and impacts; CDR-based studies measure post-event human mobility and resettlement. Taken together, they represent complementary layers of disaster evidence: one centered on event cataloguing and financial consequences, the other on behavioral recovery trajectories. A plausible implication is that disaster databases and mobility analytics can be integrated in future comparative work on post-disaster recovery.

6. Acronym ambiguity and distinction from the “Comprehensive Disaster Dataset”

In recent literature, the acronym CDD is not unique. A 2026 computer-vision paper uses CDD to denote the Comprehensive Disaster Dataset, an aerial image benchmark containing 13,316 images across five classes: infrastructure damages, fire, water disasters, human damages, and normal (Ferdaus et al., 19 Mar 2026).

That image dataset is unrelated to the Canadian Disaster Database maintained by Public Safety Canada. Its purpose is embedded image classification under resource constraints, and it is used to evaluate VeloxNet, a lightweight CNN with gMLP blocks and a Spatial Gating Unit. The paper reports 6,737 training, 1,542 validation, and 5,037 testing images, with classes heavily imbalanced toward normal, and gives a headline result of 77.46% weighted F1 on that benchmark for VeloxNet.

The acronym overlap can create ambiguity in interdisciplinary searches, particularly when disaster research spans hazard databases, mobility analysis, and computer vision. The Canadian Disaster Database is an event archive and risk-modeling substrate; the Comprehensive Disaster Dataset is an aerial imagery benchmark. Treating them as interchangeable would conflate fundamentally different data modalities, sampling frames, and inferential targets.