Unadapted

Methodology, version 2.3

How the numbers are worked out

For each event Unadapted shows what that one event cost, how often such events happen at that place and what they have cost before, what adapting would cost, and how much adapting now would save over the next 10 years and over the measures' lifetimes. This page explains each step, the formulas, and where the numbers are weakest.

What counts as an event

A specific episode of a hazard that climate change makes more likely or more severe, which has already happened and caused harm: heatwaves, floods and extreme rainfall (including rain-triggered landslides), wildfires, droughts and weather-driven crop failures, tropical cyclones and severe storms, storm surges and coastal erosion, glacier and permafrost hazards, and ecological impacts such as mass coral bleaching.

Not included: earthquakes, volcanoes and tsunamis; industrial and transport accidents; forecasts or warnings with no impact yet; and research, policy or opinion pieces that don't describe one specific event. A cyclone and the floods and landslides it causes count as one event.

How events are found

  1. Once a day, news searches across each hazard find new articles from the past two days. Syndicated copies of the same story are skipped.
  2. An AI model reads a factual summary of each article and decides whether it reports a qualifying event, extracting the place, dates and toll.
  3. Only major events are recorded: 10 or more deaths, 100,000 or more people affected, losses over $100 million, or national or international news. Articles about the same event are grouped together, by the model and by matching country, dates and place names. A smaller story can still add coverage to an event already recorded, and an event that grows into a major one is picked up from the next day's coverage.
  4. Each event is then analysed against numbered sources: its own coverage, reports of losses, the place's history and existing protection, and examples of adaptation elsewhere. A major event whose cost was our own estimate is analysed once more three weeks later, in case a published figure has appeared. Every version is kept.

The cost of an event

The total economic cost: damage to homes, property and infrastructure, losses to farming and business, and the cost of the emergency response. Where a government, insurer, reinsurer or the World Bank has published a total, we use it and mark it reported. Otherwise we build an estimate from the parts and mark it estimated. Every cost has a low, central and high figure.

All money is in US dollars at 2026 prices. Other currencies are converted at the rate of the time, and older figures are adjusted for inflation. Deaths are reported separately and are not given a money value.

How often it happens

Climate-driven hazards recur. Some towns flood every few years; parts of Spain's Mediterranean coast are hit by cut-off-low ("DANA") storms most autumns. Comparing the cost of adapting with a single event understates what adaptation is worth, because a flood wall or a warning system protects against every repeat over its life. So each event is placed in a hotspot: the hazard at the place one set of measures would protect, such as river flooding in one town. Events in the same hotspot are linked on the site.

For each hotspot we record:

The cost of adapting

The package of measures appropriate for that place and hazard, sized for the affected area and priced at local costs, with a real precedent where one exists. For each measure:

Which measures count. The cost of adapting counts only measures that pay for themselves over their lifetime on the losses they avoid (a return of at least 1 on their own), plus warning and health measures where the analysis finds lives at stake. Measures that would cost more than they save are still listed on the event page, marked as not counted, because a budget wouldn't fund them on these grounds. Where nothing pays back, the site says so rather than quoting a cost.

Measures are judged on their own, and overlapping measures that protect the same assets in the same way are not both counted. Two further shares capture the limits of any design: the share of this event's losses that were beyond what the measures are built for, and the share of expected annual losses that come from events that big.

How they compare

The model supplies the inputs above. The arithmetic is done in code, the same way for every event:

Combined effectiveness = 1 − (1 − e₁)(1 − e₂)…, capped at 95%

Each measure acts on the losses the others leave, so effectiveness doesn't simply add up, and no package avoids every loss.

Annualised cost = Σ (up-front cost × CRF) + Σ running costs

CRF, the capital recovery factor, spreads an up-front cost over the measure's life as equal yearly amounts: r(1 + r)ⁿ ÷ ((1 + r)ⁿ − 1), with lifetime n in years and a discount rate r of 3% a year in real terms. That sits between the 2% in the US government's Circular A-4 (2023) and the 3.5% in the UK Treasury's Green Book.

Losses avoided a year = expected annual loss × combined effectiveness × (1 − share beyond design)

Lifetime return (benefit-cost ratio) = losses avoided a year ÷ annualised cost

Comparing yearly amounts like this is equivalent to comparing present values over the measures' lives. A ratio above 1 means the measures save more than they cost. The low and high ratios use the low and high expected annual loss.

Each measure on its own: losses avoided a year = expected annual loss × its effectiveness × (1 − share beyond design), divided by its own annualised cost

A package can return less than its best measure: a cheap warning system may pay back many times over while an expensive scheme alongside it doesn't. We show both, and call out the best-value measure when it clearly beats the package.

Payback = up-front cost ÷ (losses avoided a year − running costs), for the measures that count; "never" if running costs eat all the savings

Saves in 10 years = losses avoided a year, summed over the next 10 years

Cost over 10 years = up-front cost + 10 years of running costs; the difference between the two is shown in dollars

The forward view a budget-holder would ask for: if the measures were built now, what would the next 10 years look like? Losses avoided are the losses avoided a year, summed over 10 years; the cost is the full up-front cost plus 10 years of running costs. Both are in today's dollars, discounted at 3%. Counting the whole up-front cost against only 10 years makes this a demanding test for long-lived defences, which is why the lifetime return is shown too.

Looking back = expected annual loss × 10 years, and the losses avoided a year × 10

What not adapting has cost over the last 10 years, in 2026 dollars, at the place's expected rate; most earlier events have no published cost, so a sum of published figures alone would understate it. Those published figures are shown alongside.

This event alone: losses avoided = event cost × combined effectiveness × (1 − this event's share beyond design)

A worked example

A town floods every few years, with expected losses of $6M a year. The latest flood cost $30M. The package is flood walls ($40M up front, $200K a year, 50 years, avoiding 80%) and a warning system ($1M, $100K a year, 10 years, avoiding 25%). 10% of expected losses come from floods bigger than the walls are built for.

Judged against the latest flood alone, the same package looks poor value: it would have avoided about $25.5M against $41M up front. That is why the lifetime return is the headline figure.

Is it climate change?

Each event carries one of three levels of evidence:

"Fits the warming trend" relies on the IPCC's Sixth Assessment Report, which finds that some hazards, such as hot extremes, are becoming more frequent and intense almost everywhere, while the evidence for others, such as heavy rainfall in some regions, is weaker.

The cost shown is the full cost of the event, not just the part caused by climate change. Where a formal attribution study gives a probability ratio (PR: climate change made the event PR times more likely), we also show the climate-attributable share using the fraction of attributable risk, 1 − 1/PR (Stott and others, 2004), applied to the cost as in Newman and Noy (2023). For example, an event made twice as likely has half its risk, and so roughly half its cost, attributed to climate change. Adaptation pays off whatever caused the event, so the lifetime return doesn't depend on attribution.

Existing protection

Each event records what protection was in place or planned beforehand:

Lives and homes

Deaths are reported, not priced, and are left out of the cost comparison. That makes the lifetime return conservative for hazards where the main harm is loss of life, such as heatwaves. Where there is evidence that the measures save lives (for example heat-health warning systems or flood warnings), we show an estimate of the lives they would likely have saved, with its basis. For hazards that damage homes (floods, storms, fires, coastal surges), we also show how many of the homes damaged the measures would likely have protected.

Uncertainty

Costs, expected losses and returns come as low, central and high figures, and each analysis carries a confidence level (low, medium or high) with its key assumptions. Reported costs for recent large events are usually the most reliable figures. Expected annual losses, effectiveness and costs for places with thin records are the least reliable.

Edge cases

Limitations

AI and review

Articles are found with Exa's search. Screening and analysis are done by Anthropic's Claude Opus 5.5 working from the numbered sources on each event page. Analyses are published automatically and are not reviewed by a person before they appear. Every figure links to the sources it relies on, so it can be checked.

Changes

References