US Extreme Weather and Climate Change Dashboard

Methodology

This entire site is possible because of publicly available data, funded by taxpayers and governments around the world. Open science is good science.

Detection and Attribution

This dashboard focuses only on detection, following the IPCC's own framework for detecting a change in climate. The IPCC's definitions are below (Glossary, AR5/AR6/SR15, "Detection and Attribution"), and are applied throughout this site.

• Climate: "The average weather, or more rigorously, the statistical description in terms of the mean and variability of relevant quantities over a period of time ranging from months to thousands or millions of years."
• Climate change: "A change in the state of the climate that can be identified (e.g., by using statistical tests) by changes in the mean and/or the variability of its properties, and that persists for an extended period, typically decades or longer."
• Detection: "The process of demonstrating that climate or a system affected by climate has changed in some defined statistical sense, without providing a reason for that change. An identified change is detected in observations if its likelihood of occurrence by chance due to internal variability alone is determined to be small, for example, <10%."
• Attribution: "The process of evaluating the relative contributions of multiple causal factors to a change or event with a formal assessment of confidence."

Detection and attribution are two distinct steps, and this dashboard performs only the first. Detection asks whether observations have changed in a defined statistical sense, without saying why. Attribution asks why — evaluating the relative contribution of specific causal factors (greenhouse gas forcing, land-use change, natural variability, and so on) typically requires climate model simulations run with and without a given forcing, a substantially larger undertaking this project has not attempted. Nothing on this site is an attribution finding: a panel marked "Detected Change" means the observed trend is unlikely to be due to chance internal variability and is large relative to that variability — not that a cause, human or otherwise, has been established.

Detecting a change is also not the same as spotting a trend in a chart. Any naturally variable quantity drifts up and down over any finite stretch purely from year-to-year (or week-to-week) variability, with no underlying change in the climate at all — a trend line by itself proves nothing. IPCC's detection criterion asks whether that trend is unlikely to have arisen from natural internal variability by chance alone, specifically at the stated example threshold of below 10% (via the nonparametric Mann-Kendall test used throughout this site). This dashboard goes one step beyond IPCC's minimum: because a long, low-noise record can register a statistically real trend even when its actual size is trivial next to the variable's natural variability, a trend is only called adetected change here if it also clears a second, magnitude-based bar — see "Magnitude-vs-variability check" below. A trend can clear IPCC's likelihood criterion and still not count as a detected change on this site for that reason.

Sources

• Heat waves: reconstructed directly from NOAA's GHCN-Daily station records (the ~1,218-station HCN reference network), since NOAA's named source for this variable — the Accumulated Heat Wave Index — turned out to be a frozen, one-off analysis that stopped updating in 2021, not the live product it appeared to be. A per-station 90th-percentile trailing-4-day-mean threshold stands in for Kunkel et al.'s original (not fully public) method.
• Tornadoes & hail: NCEI's Storm Events Database, live and continuously updated. Storm Events records one row per county a tornado's path crosses rather than one row per tornado — physical tornado tracks are reconstructed from those segments before counting (see Known Caveats).
• Streamflow: USGS's HCDN-2009/GAGES-II reference-gauge network (741 gauges, minimally altered by dams/diversions/urbanization), queried live against USGS's National Water Information System.
• Drought: the US Drought Monitor's weekly categorical area percentages (2000–present) as the headline series, with NOAA nClimDiv's national Palmer Drought Severity Index (1895–present) as a long-record context layer.
• Winter storms: NCEI's Regional Snowfall Index storm catalog, accessed via its live ArcGIS REST service (its landing page has no working download link — the actual live data lives one layer down).
• Wildfire: MTBS (Monitoring Trends in Burned Severity, satellite-measured burned area, 1984–present) and NIFC (National Interagency Fire Center, fire count, 1983–present) — two independent products with different methodologies, shown as separate panels rather than merged into one series.
• US landfalling hurricanes: NHC's HURDAT2 Atlantic best-track archive, downloaded and parsed directly. Landfalls are detected from each storm's fix coordinates against a real US country boundary, not from any single source's landfall flag (HURDAT2's flag has real, documented gaps — see Known Caveats).

Definitions

• Reliable-trend window: every variable has a longer raw record than the window over which its trend is meaningfully comparable year to year — instrumentation and methodology changes (Doppler radar deployment, satellite-based measurement, reference-network construction) mean the earliest years often aren't measuring the same thing as recent ones. Each panel defaults to a documented window but can be widened or narrowed with the window-start/window-end controls.
• Mann-Kendall test & Sen's slope: the trend statistics used throughout — nonparametric methods that don't assume normally-distributed data (several of these series are bounded, skewed, or zero-heavy), standard in the climate-trend literature for exactly this kind of question. Reported alongside a two-sided p-value; a trend is called "statistically significant" only at p < 0.10 (see "Detected Change" below).
• "Detected Change": this site's combined standard, applied throughout — see "Detection and Attribution" above for the underlying IPCC definitions. A trend counts as a detected change only if it clears both: (1) IPCC's likelihood criterion, treating the Mann-Kendall p-value as that likelihood (p < 0.10, per IPCC's "for example, <10%" language — deliberately looser than the common p < 0.05 convention, because that's the threshold IPCC's language actually specifies), and (2) the magnitude check below — since a likelihood pass alone doesn't say whether the change is large. A trend that clears only the likelihood criterion is reported as statistically real but labeled "No Detected Change" rather than a qualified "detected."
• Magnitude-vs-variability check: IPCC's glossary text is a likelihood definition and doesn't itself specify a magnitude threshold, so this third check is this project's disclosed judgment call, not an IPCC-mandated number. Computed as the trend's total fitted change (Sen's slope × window length) as a share of the variable's 90% ("very likely") historical range -- the same band shown as shading on each chart. A trend needs at least 25% of that range to count toward a detected change — found directly in this site's data as the reason a statistically real trend can still read "No Detected Change": streamflow's trend is statistically real (p=0.011) but only ~4% of its 90% historical range, vs. wildfire's MTBS area-burned trend (also statistically real, p < 0.001) at ~63% of its range.
• 66% / 90% shading: IPCC's calibrated-language convention ("likely" / "very likely") applied to each variable's historical distribution within the selected window — the 17th–83rd and 5th–95th percentile range of observed values, shown as a reference envelope for how much this variable normally varies, not a forecast interval.
• Saffir-Simpson category / F-EF scale: hurricanes and tornadoes are both binned by NHC's and NWS's standard severity scales respectively, assigned from the recorded wind at the specific event (landfall fix for hurricanes, damage survey for tornadoes) rather than a storm's peak intensity at any other point in its life.

Known caveats

• Every finding on this site is CONUS-specific, not a global claim. This dashboard's whole scope is US weather and climate data — no panel here says anything about global trends, and a US pattern (a record year, a detected change, a "Dataset Dependent" disagreement) can and often does run opposite to the global picture over the same period. The 1930s Dust Bowl is the clearest example: it's a real, well-documented CONUS-scale heat extreme visible in this site's own heat-wave data, but global temperatures over the same years were cooler than today — treat any US-only finding here as exactly that, not evidence about the planet as a whole.
• Heat waves' headline detection metric is WSDI computed on NOAA's homogenized, gridded nClimGrid-Daily product, deliberately chosen over this site's own raw per-station panels (heat wave index, TXx, WSDI, TN90p — all an unweighted mean across raw, non-homogenized GHCN-Daily stations, with no correction for instrument changes, station moves, or time-of-observation shifts, and no adjustment for uneven station density) because trend detection specifically calls for homogenized data, and because it's the same standard the IPCC/NCA literature itself uses. That raw-station combination is exactly the kind of pitfall that has produced misleading "hotter in the 1930s" claims elsewhere — this site's own raw per-station WSDI shows a small, real decline since 1895 (not a detected change once weighed against variability), a genuinely different picture from the homogenized headline panel's clear detected increase since 1951. Both are shown, side by side, rather than merged into one number; an independent, peer-reviewed analysis using similarly raw, unadjusted station data reaches a comparable raw-data conclusion (Christy 2026). Where this site's own metrics disagree with each other, the summary tables say "Dataset Dependent" instead of implying consensus.
• Hail has no reliable-trend window at any point in its record — report counts track population density, spotter-network growth, and reporting practice far more than any real climate signal, a limitation NCA5 (2023) states explicitly. Shown with a permanent, prominent caveat rather than a trend line.
• Streamflow and USDM drought's windows are download-boundary artifacts, not discovered data-quality boundaries — unlike heat waves' empirically-found 1895 cutoff, these two variables' start years (1990 and 2000 respectively) are simply how far back each pipeline's live feed reaches, so their whole downloaded record is treated as the reliable window.
• Winter storms uses Category 3+ (major) counts, not the raw storm count — the raw count likely conflates a real climate signal with growth in the observing network over the last century.
• Wildfire's two panels (MTBS area burned, NIFC fire count) measure different things on different timelines and shouldn't be read as a single series — MTBS's latest year typically lags NIFC's by 1–2 years, and NIFC's live data doesn't extend before 1983 at all (contradicting some older published tables that imply a longer record).
• Hurricane landfall counts use a simplified continental-US geometry test (a single buffered country polygon plus a bounding box, not a full coastline/land-mask reconstruction), and a storm that weakens below hurricane intensity before actually reaching the coast is correctly excluded even if it's popularly remembered as a hurricane (e.g. Fern, 1971). See the Methodology & Sources PDF for the full validation against the known Category-5 landfall list.
• A side-by-side comparison of how IPCC AR6 has characterized each hazard, and how it compares to this site's detected-change findings, lives on the Detection & Attribution page. A fuller three-assessment comparison (adding NCA4/CSSR 2018 and NCA5 2023) is a possible future expansion, not yet built.

Full per-variable detail, including validation against known historical events and every documented limitation, is available in the Methodology & Sources PDF.