By Ron Dembo
There are two ways to predict flood damage in a warming world. One adjusts the past. The other rebuilds from the weather up. They start from the same climate data and end at the same flood simulation. They differ at exactly one step — how warming gets into the model — and that step decides everything downstream.
This isn't a debate about whose flood physics is better. Both approaches can produce detailed, defensible-looking output at a single location. The difference shows up somewhere else entirely: in whether the model can answer the questions that actually move a balance sheet.
Most flood models in commercial use today start with a fixed list of historical flood events and nudge them worse under warming — take last year's weather and add a percentage for climate change. The relationship they assume between future weather and past weather is linear: a fixed adjustment, stretched across the record, applied the same way whether the projection reaches five years out or fifty. It's practical. The flood detail at a given location can be excellent. But the warming is bolted on, not rebuilt. The model learns a rule from history and assumes the rule — and its straight-line shape — still holds in a much warmer world.
That assumption holds for average conditions. It breaks for the extremes that determine tail losses.
The alternative doesn't extrapolate a single adjustment across the future. It resimulates the atmosphere at each future point, so floods, droughts, heat, and storms all emerge freshly from the physics of that point in time rather than from a percentage applied to the past. More computation. Necessary, because the relationship between warming and weather doesn't stay constant — it changes shape as the climate moves further from the record the old models were trained on. Because every hazard comes from a single shared weather picture, the model can also show how hazards occur together — a flood on top of a drought-baked soil, a storm surge arriving with river flooding from the same system. That's not an added feature. It's a structural consequence of how the model is built.
A warmer atmosphere doesn't just shift the average — it changes the mechanics. Dry, baked soil sheds water instead of absorbing it. Storms shift track and shape over decades. A 1-in-100-year storm in 1990 is not the same event as a 1-in-100-year storm in 2090, and a model trained on the historical record has no way to know that.
The worst losses rarely come from a single flood. A drought-baked region hit by heavy rain can turn a 1-in-100-year rainfall event into a 1-in-500-year flood. Storm surge, river flooding, and rainfall, treated as unrelated hazards, look rare individually. Because they share one storm system, they strike together far more often than independent-hazard models suggest — and the information needed to fix this was discarded early in the "adjust the past" pipeline. It can't be recovered by refining the model further.
Old models assume every flood arrives at a fully recovered town. In practice: soil still saturated, pumps still broken, half-repaired homes destroyed again before the first repair is finished.
The 2017 hurricane season — three major storms in six weeks — caused roughly 40% more damage than three separate storms would have produced. The 2022 Pakistan floods caused close to three times the damage that simple, single-event models predicted.
Take a regional bank holding roughly $4 billion in commercial property loans across the Gulf Coast and Mid-Atlantic, most maturing in three to five years.
Run the portfolio through an adjust-the-past model: expected losses rise only slightly by 2050, capital requirements barely move, and the recommendation is to hold the book and monitor annually. The model isn't wrong — it did exactly what it was built to do.
Run the same portfolio through a rebuild-from-weather-up model, and three things change. Hazards stack: one hurricane can hit storm surge, river flooding, and rainfall across many properties simultaneously, rather than as separate low-probability events. Repeat hits show up: the three-to-five-year loan window is exactly where a repeat of the 2017 hurricane cluster would land, pushing worst-case loss toward 10–15× rather than the 5× simple math assumes. And the worst-case number itself typically moves 2 to 5 times higher.
The average loss barely changes between the two models. The worst-case loss — the number capital rules and stress tests actually depend on — moves substantially. Both models are correct for what they were built to do. Only one of them answers the question the capital decision turns on.
Six requirements define a model built to answer the hard questions, consistent with the Climate Risk 2.0 standard:
The frontier of flood research has already moved past adjust-the-past methods, toward approaches that treat repeat events and stacked disasters as core features rather than afterthoughts. Those methods aren't yet commercial-grade — they're expensive to run and harder to validate. But a rebuild-from-weather-up architecture is designed to absorb them as they mature. An adjust-the-past model would need to be torn down and rebuilt from scratch to do the same.
Adjust-the-past models will keep getting better. They'll just be better versions of a model built on a linear assumption — one adjustment, stretched across a future that doesn't move in a straight line. They were never built to answer the questions with real money behind them: multiple disasters at once, repeated events, combined extreme losses, long-lived assets.
The question worth asking any flood model vendor isn't which one has the best flood physics at a single point. It's which one resimulates the atmosphere at each future point instead of extrapolating from the last one — and which can keep doing that as the science moves forward. Our CDTexpress engine is built on the rebuild-from-weather-up approach for exactly that reason — the same architecture behind the Climate Risk 2.0 standard, applied to flood.
Access the full report here