Why Mankind Will Not Disappear Due to Wars: Pierre Rostan Projects 976 Years of Conflict Deaths

In the year 2238, a forecasting model built by two researchers at the American University of Iraq – Baghdad puts the global death toll from war at 1,957,513. That is the worst single year the model produces for the rest of this millennium. It is also less than a sixth of the deaths recorded in 1939.

The distance between those two numbers is the argument of a paper by Pierre Rostan, an associate professor in the College of Business, and Alexandra Rostan, a researcher there, published in Conflict Resolution Quarterly. Across the 976 years from 2025 to 3000, they estimate, wars will kill an average of 843,882 people annually. That is 4.7 times the annual average of 177,186 the authors calculate for the 193 wars they catalogued between 549 BC and 2024. Measured against world population, though, the death toll never reaches 0.016% in any year, peaking in 2071. The historical maximum, 0.977%, falls in 755–758, during the An Lushan rebellion in China.

One person in roughly 6,250, at the worst moment of the next thousand years. That is the arithmetic behind the paper's title.

Timing matters here. The Uppsala Conflict Data Program recorded 61 active state-based conflicts in 2024, the highest count since its records began in 1946, and nearly 160,000 deaths from organized violence. A projection saying war stays survivable in percentage terms is easy to misread as a projection saying war stays small. The Rostans don't argue that. Their own figures put the annual toll well above anything in the post-1945 record.

What the model does, stripped of jargon

The input is a list. The authors assembled death-toll estimates for 193 wars from 549 BC to 2024 — 456,078,636 casualties spread across 2,574 years — and turned that sequence into a time series, a single line of numbers running through history.

Then they borrowed a tool from signal processing. Think of a sound engineer handed one long recording of an orchestra and asked what happens next. She can't simply extrapolate the whole mush forward. So she separates the bass line from the strings and the cymbals, examines each layer's rhythm, and extends each one on its own before mixing them back together. Wavelet analysis does that to a data series: it splits a signal into slow-moving approximations and fast-moving details, each at its own scale.

The technique is the same one Rostan has applied to GDP growth, pension sustainability and fossil fuel prices in earlier work. Here the four steps run like this:

• The series is differenced, de-noised and compressed, stripping out year-to-year noise so the underlying shape shows through.
• Wavelet decomposition breaks that shape into approximation components, which carry the long trend, and detail components, which carry the cycles and shocks.
• A Burg model extends each component. It's an autoregressive predictor that fits coefficients by minimising forward and backward prediction errors at once, then projects them outward.
• Wavelet reconstruction reassembles the extended pieces into a forecast series running to the year 3000.

The resulting death tolls are then divided by United Nations-style projections of global population, which is where the percentages come from. That division is the hinge of the whole paper, and it's worth being clear about what it does and doesn't measure.

The numbers, as the paper reports them

Annual deaths climb through the 21st century and peak in 2238. During the third millennium the model never produces a year above 2 million deaths. A second peak arrives in 2483 at 1,871,680, after which the curve turns down, reaching 660,809 in the year 3000.

Two of those rows pull in opposite directions, and the tension is the interesting part. The average annual toll nearly quintuples compared with the historical record. The cumulative toll over the next 976 years comes out 2.5 times larger than the cumulative toll of the previous 976. Yet the death rate as a fraction of humanity falls by more than an order of magnitude, because the denominator grows so much faster than the numerator.

Photo by Saifee Art on Unsplash

Why the percentage falls while the deaths rise

Population is doing most of the work. The United Nations projects world population peaking around 10.3 billion in the mid-2080s and easing to about 10.2 billion by 2100. Over the following centuries the Rostans' comparison assumes a global population large enough that even two million war deaths in a single year remain a rounding error in percentage terms.

Their benchmark for what a genuinely catastrophic war looks like as a share of humanity is the An Lushan rebellion. That eighth-century Chinese uprising is conventionally assigned millions of deaths against a much smaller world population, and it produces the 0.977% figure. World War II, which killed 12,047,636 people in 1939 alone, was proportionally less lethal because the world held roughly 2.3 billion people by its end.

Put plainly: industrial total war was the outlier, not the trend line. A model trained on 2,574 years of data treats 1939 as a spike and reverts toward a lower mean. That is a defensible statistical choice and also the paper's largest point of exposure. The full study is available through its DOI record at Wiley.

Where the forecast is most exposed

The first vulnerability is the input data. Pre-modern death tolls are estimates built from tax records, chronicles and population declines, and historians argue about them constantly. The An Lushan figure, cited in the paper's own comparison, has been estimated at 13 million by some scholars; Pasquale Cirillo and Nassim Nicholas Taleb note in their study of war casualty statistics that the true number could conceivably be an order of magnitude smaller. If the benchmark disasters are softer than they look, the ratio the paper leans on moves too.

The second is the shape of the distribution. War deaths follow a fat-tailed pattern: rare, enormous events dominate the total. Cirillo and Taleb applied extreme value theory to roughly 2,000 years of conflict data and concluded that conventional analyses systematically underestimate the risk of the largest wars, and that the arrival of big wars shows no discernible trend. A method that extrapolates cycles from observed history has limited purchase on an event that has never been observed.

The third is nuclear weapons, which appear nowhere in a series ending in 2024. Climate and crop modelling by Lili Xia, Alan Robock and colleagues, published in Nature Food, estimates that more than five billion people could die of hunger after a full-scale war between the United States and Russia, with soot from firestorms cutting global calorie production by around 90% within four years. No curve fitted to 549 BC–2024 can carry that scenario. The model's ceiling of two million deaths in any single year is a statement about the world it was trained on, not a guarantee about the one that exists.

A fourth limitation is structural: the forecast depends on population projections reaching eight centuries beyond the UN's own horizon of 2100. Small differences in assumed fertility compound enormously at that distance.

What the authors are and aren't claiming

The paper's question is narrow and deliberately so. It asks whether wars can eliminate humanity, and answers no, on the arithmetic. It does not ask whether war will become less frequent, less cruel, or less concentrated among civilians. The Rostans' own projection has the annual toll rising to nearly two million, roughly eleven times the annual average of the preceding two and a half millennia, and staying in that range for centuries.

That distinction matters for how the finding gets used. A politician could read "mankind will not disappear" as permission to treat war as a manageable background condition. The model doesn't support that reading any more than it supports the opposite one. It says the species survives; it says nothing about whether the century ahead is one in which conflict resolution succeeds or fails.

Where the paper is genuinely useful is as a scale check. Numbers in the hundreds of thousands per year are large enough to demand attention and small enough to be lost in coverage of a single dramatic conflict. Setting a projected 843,882 annual deaths beside a projected 0.016% share makes both figures visible at once.

Photo by Evgenii Vasilenko on Unsplash

What to watch over the next decade

The near-term forecasts are the ones worth testing. The model's peak in proportional terms falls in 2071, which means the next fifty years are where the ratio is meant to climb. If the 2020s and 2030s produce annual tolls far above the projected path toward 2238, the fitted cycles are wrong and the paper's authors would be the first to want the discrepancy on record.

Meanwhile the underlying question — whether a long historical record can say anything useful about the next thousand years — is being asked across disciplines, from pandemic forecasting to climate economics. Rostan's answer is that the cycles in the data are real enough to extend. The sceptics' answer is that the biggest events are precisely the ones with no precedent to extend from. Both are correct, which is why the paper is best read as a floor rather than a forecast: humanity survives wars of the size it has already seen, and the historical record is silent on the ones it hasn't.