Polycrisis Resilience This Century, 2000–2100

The world death rate from disasters averaged 9.0 per 100,000 in the 1900s, 26.5 in the 1920s (the great droughts and famines in China and the Soviet Union), 10.2 in the 1960s, 2.5 in the 1970s, 1.6 in the 1980s and 0.77 in the 1990s: a fall of 97% from the 1920s peak. The first decades of the twenty-first century are flat: 1.18 in the 2000s, 0.64 in the 2010s and 0.74 in 2020–25. The median year is more striking: 0.43 in the 2000s, 0.16 in the 2010s and 0.72 in 2020–25. In the 2010s a typical year was quiet; in the 2020s it is not.

Catastrophe has not receded. Years with 100,000 or more deaths worldwide were 26 of 100 in the twentieth century (28% in 1900–49, 24% in 1950–99) and 5 of 26 so far this century (19%): 2003, 2004, 2008, 2010 and 2023. What has changed is the cause. Years with a million or more deaths, six of them before 1950, five from drought and famine and one (1931) from the China floods, have not occurred since. The large years are now earthquakes (Indian Ocean 2004, Haiti 2010, Türkiye and Syria 2023), cyclones (Myanmar 2008) and heatwaves (Europe 2003). Meanwhile heat's share of recorded deaths rose from 3% in 2005–09 and 16–20% in the 2000s and 2010s to 58% in 2020–25, as earthquakes fell from 64% to 21%.

Two cautions. EM-DAT's coverage improves over time, so early decades are undercounts and the earlier decline is, if anything, understated. Heat deaths are the least reliably recorded: they are often estimated from excess mortality after the event, and recent heatwaves have been studied more closely than earlier ones, so part of heat's rise is better counting. And recorded damage as a share of GDP shows no trend, at about 0.2% a year in five-year blocks. The plateau is in deaths and in damage alike.

The 169 countries (8.2 billion people in 2025, 10.1 billion in 2100 on the UN's medium variant) change composition sharply. Sub-Saharan Africa goes from 1.3 billion in 2025 to 3.5 billion in 2100, its share of this population from 16% to 34%. The Middle East and North Africa rises from 10% to 15%. East Asia falls from 2.3 billion (29%) to 1.5 billion (15%), Europe and Central Asia from 11% to 8%, South Asia from 21% to 17.5% after peaking near 2.0 billion in 2065.

At 2000–2024 death rates the expected deaths move with the people, but less than proportionally, because Africa's recorded disaster mortality is low. East Asia accounts for 32% of expected deaths in 2025–50 and 27% in 2051–2100; Europe and Central Asia 20% then 16%; Latin America 18% then 19%; the Middle East and North Africa 13.5% then 17%; South Asia 10% in both; and sub-Saharan Africa 6% then 10%, against its 34% share of people. The expected deaths are dominated by where the great events of 2000–24 happened (Indonesia, Haiti, Myanmar, China, Pakistan).

That is a reason for caution, not comfort. Recorded deaths in sub-Saharan Africa are low in part because EM-DAT includes an event only above a size threshold and has weak coverage of slow-onset and heat events in low-income settings; because the continent's population is young; and because the 25-year record may not include its own catastrophes. A drought and a heatwave on a population of 3.5 billion would not look like the record. The simulation scales deaths with people; it cannot tell where a hazard will fall.

Each scenario multiplies 2000–24 death rates by (1 − capacity gain) × (1 + hazard growth) for every year after 2024. A (frozen) is the record meeting the population: the expected annual deaths rise from about 38,000 in the median year of 2025 to 42,000 in the median year of 2100 and a mean of 81,000, because the mean is carried by the big years. B (capacity +1% a year) takes the pace at which the mean world rate has fallen since the 1990s (0.77 to 0.64 per 100,000 in the 2010s, about 1% a year) and halves the median 2100 year to 19,000. C (stall and warm) lets hazards grow 1% a year with no capacity gain, so rates more than double by 2100 and the median year reaches 89,000, with a mean of 179,000. D (capacity push) improves capacity 3% a year against the same hazard growth, a net decline of 2% a year, and the median 2100 year falls to 9,000.

The 5–95% range of the century's total within a scenario is about ±25% (4.9–7.7 million for A), because 169 countries and 76 years average out much of the luck. The range between scenarios is a factor of three (3.0 million for D to 9.3 for C). Of the two, the trend is the bigger source of uncertainty: whether capacity improves and whether hazards grow are not statistical questions, and the record cannot settle them. The hazard assumption is a parameter, not an estimate of climate-driven mortality; the report does not claim to know what growth of hazard is likely, and the calculator lets the reader choose.

The leverage is symmetrical and large. From frozen rates, each percentage point of annual capacity gain removes about 2 million deaths over the century and each point of annual hazard growth adds about 3 million (because it compounds on a growing population). A world that holds capacity growth equal to hazard growth stays about where it is: 6.2 million deaths for A, and for a 1%-and-1% world also about 6 million. The difference between a good and a bad century is about 6 million lives, and it is decided by policy and climate, not by chance.

The frequency of 100,000-death years in the simulation, about one in five, matches the record: 26% in the twentieth century, 22% since 1950, 19% since 2000. That is a validation of the simulation's overall tail, not of its geography. At frozen rates the simulation expects 15 such years in 2025–2100; B gives 12, C 25 and D 8. For years of 250,000 or more, A and C reach one in virtually every simulated century (B 92%, D 64%); for 500,000 or more the chance is 28% under A, 3% under B, 96% under C and 1% under D; and for a million or more only C, at 12%, has a material chance.

These probabilities are what insurers and finance ministries need. A government or a humanitarian system that plans for the average year (about 80,000 deaths a year at frozen rates, 179,000 by 2100 under C) will be unprepared for the median worst year, which is 434,000 at frozen rates and 764,000 under C. Preparing for a year with several hundred thousand deaths, with reserves, pre-agreed finance, surge capacity and the logistics of response, is a standing requirement of the century, not a contingency.

Two limitations bias the tails in opposite directions. The simulation draws each country's years from its own 25-year record, so a country with one catastrophe repeats it at about 4% a year, which overstates the frequency of single-country events (it expects 12 country-years above 100,000 deaths a century, against three in 2000–2024) and understates catastrophes in countries that happened to have none. And it treats countries as independent, so it understates correlated years in which one cause (a heat dome, a monsoon failure) strikes many countries. Neither affects the world-level frequency, which matches the record, but both mean the geography of the tails should not be read literally.

The scenarios diverge slowly and then quickly. In 2045–54 the mean annual deaths are 48,000 under the capacity push and 106,000 under stall-and-warm; in 2095–2100 they are 18,000 and 171,000 (at frozen rates they stay near 82,000). Most of the 6 million lives between the best and worst paths are therefore in the second half of the century, which is why a policy that looks costly now has benefits that will be realised after most of today's budget cycles. Compound growth, not a single investment, produces the gap.

Relative to frozen rates, capacity improving 1% a year saves about 1.9 million lives over the century, and the capacity push, even with hazards growing 1% a year, saves 3.2 million; hazards growing 1% a year with no capacity response cost an extra 3.2 million. The mean at frozen rates is already 4.0 times the 1.55 million recorded in 2000–2024, because the world adds people and the simulation repeats the record's catastrophes; it describes a bigger, more exposed world, not a worse climate.

Who pays follows the population clock of section 2: more of the deaths fall in the Middle East, North Africa and sub-Saharan Africa as their populations rise, and more in Latin America, where recorded rates are high. Targeting capacity on the regions where exposure grows, and on heat, which is the fastest-growing recorded cause, is consistent with these results; so is a high priority for recording: the places where the simulation expects few deaths are in part the places where deaths are least well counted.

The simulator draws a year from each country's own 2000–2024 record of death rates for every year from 2025 to the horizon, multiplies the rate by the country's projected population and by the compounding capacity and hazard factors, and sums over the countries of the chosen region. It repeats 800 times and reports medians and percentiles. The calculator is a faster, smaller version of the 10,000-simulation runs behind the charts, so its numbers vary by a percent or two from run to run and from the static figures.

Reading the output: the median and the 5–95% range give the century's total; the average deaths a year and the median final year show the level; the expected number of years above 100,000 and the chance of a year above 500,000 describe the tail; and the difference from frozen rates shows what the chosen trends are worth. Setting capacity equal to hazard growth gives a century close to frozen rates; setting hazard to zero shows the value of capacity alone.

The simulator inherits the limits of the record: 25 years of data, a size threshold in EM-DAT, countries drawn independently and the repetition of each country's catastrophes. It uses medium-variant population projections, which are themselves uncertain by a billion or more by 2100. It does not model pandemics, conflict, food or financial crises, which are the other channels of polycrisis, and it applies capacity and hazard trends uniformly to every country and hazard.

Plan for the one-in-five year. A year with 100,000 or more disaster deaths occurs about one year in five in every period of the record and would continue to occur in every scenario. Governments, development banks and humanitarian systems should hold reserves, contingent finance and surge capacity sized to a bad year, not an average one: the median worst year of a century is 270,000 to 760,000 deaths depending on the path.

Treat heat as the new mass-casualty hazard. Extreme heat's share of recorded disaster deaths rose to 58% in 2020–25. Heat action plans, cooling for the vulnerable, workplace rules and early warning for heat are cheap relative to earthquake retrofitting, and their record is encouraging but thin. Counting heat deaths consistently is the first task.

Go where the people will be, and count them. By 2100 one in three people will live in sub-Saharan Africa, where recording is thinnest. Capacity (warning, building codes, health, social protection) must be built there ahead of the population, and national loss accounting needs to improve, so that the next century's record is not another 25 years of unknowns. The difference between a good and a bad century is six million lives; it is made by capacity growing faster than hazard, not by luck.

History: Our World in Data / EM-DAT series "Number of deaths from natural disasters" (world, 1900–2025, by hazard) and "Death rates from natural disasters" (world, 1900–2025; countries 2000–2024), and "Economic damages from natural disasters as a share of GDP" (countries 2000–2024). Country-years without a record are zero. Population: "Population with UN projections", medium variant, annual 2000–2100. The simulation covers the 169 countries with at least 0.5 million people and WPP data (Kosovo has no WPP series). Decadal figures use the mean of annual world rates; 2020–25 is six years.

Simulation: for each simulated century, each year 2025–2100 and each country, a year is drawn at random with replacement from the country's 25 observed annual death rates (per 100,000); deaths = rate ÷ 100,000 × projected population × [(1 − capacity) × (1 + hazard)]^(years since 2024); world deaths are summed over countries. Scenarios: A (0%, 0%), B (capacity 1%, hazard 0%), C (0%, hazard 1%), D (capacity 3%, hazard 1%). 10,000 simulations per scenario (calculator: 800 with a fixed seed). Percentiles are across simulations; annual percentiles across years of the same simulation set. Expected deaths by country and region are analytic: mean rate × population × factor. The decade of capacity gain in B is derived from the fall in the world's mean death rate from 0.77 (1990s) to 0.64 (2010s).

Caveats: EM-DAT includes events by threshold and its coverage grows with time; heat deaths are estimated unevenly; the country records are 25 years long and dominated by a few catastrophes, which the simulation repeats at about 4% a year each; countries are drawn independently, so correlated global years are understated; the world frequency of years above 100,000 deaths is validated against the 1950–2025 record; capacity and hazard growth are assumptions applied uniformly; population uses the medium variant; and disasters are one channel of polycrisis, so nothing here covers pandemics, conflict, food or finance.