5 real US droughts, reconstructed from nothing but soil physics

One of the most convincing checks a weather model can pass is finding real historical events it was never told to look for. Anyone can build a model that's internally consistent — that doesn't make it right.

The method

For each of the 12 covered cities, we ranked every backtest year from 1940 to 2025 purely by the model's own soil-moisture-deficit numbers, then checked the driest years against real, documented, named US droughts. The model never saw a list of drought years; it only ever computed a water balance from temperature, rainfall, and evapotranspiration.

What came back

• The Northeast/Mid-Atlantic drought of record, 1961-1966 (New England's worst on record, with real NYC and NJ water restrictions) — hit Boston (1965), New York City (1966, 1964), Washington DC (1966, 1964, 1963), and Pittsburgh (1965, 1963).
• The 1950s Midwest drought — 1953 was St. Louis's literal driest year on record — hit St. Louis (1953, 1952), Chicago (1953), and Detroit (1952, 1953).
• The 1988-1990 North American drought, one of the costliest in US history — hit Minneapolis (1988), Detroit (1988), and Pittsburgh (1988).
• The 2012 US drought, the second most widespread on record — hit five of the twelve cities: Chicago, Minneapolis, Denver, St. Louis, and Salt Lake City.
• The Great Western Snow Drought of 2015 (drought.gov's own name for it — a warm winter, low-snowpack event specific to the Pacific Northwest) — hit Seattle and Portland in the exact same year.

Every one of the 12 cities landed inside a real drought window at least once. Several hit three or four of their five driest years — Pittsburgh among them, at three of five.

How strong a signal that actually is

Across all 60 driest-year entries (12 cities × their top 5), 25 — 42% — fall inside one of these named drought windows. Those windows together span only about 14 of the 86 backtest years, roughly 16% of the record. A random match rate would land close to 16%; the model landed at 42%, close to 2.6x over-representation.

One honest caveat: several of these are short windows (3-6 years) rather than single-year exact hits — a slightly less stringent test than matching one specific year. Still a real, specific signal, not a vague "it was probably dry" impression.

What it doesn't mean

This doesn't make the model infallible — weather forecasting and biology both carry real uncertainty, and this validation work is ongoing as coverage expands. But every number here earned its place by being checked against real, dated events, not by sounding reasonable in a spreadsheet.

Read the full validation approach on how it works, or sign up and let it do the watching for your own city.