How this site works
We collect reports from official agencies. We make two kinds of forecast: a few made days in advance by our own weather models, and many that restate an alert an agency has already issued. Later, we check each forecast against what really happened, each by its own rule. We never claim to be certain. You can see how our forecasts turned out on the Track record page.
1 · We collect reports
We read current events and warnings from public sources. These include the GDACS, the USGS, the EMSC, NASA EONET, ReliefWeb and US weather and hurricane services. Each source covers only some hazards and places. If no source reports on a place, we cannot see what is happening there. A place with no reports is not necessarily safe.
2 · We tidy them
We give every event the same basic details: the hazard, the time, the place and the source. When two agencies report the same event, we show it once.
3 · We sort them
You can sort events by date, risk level, place or type. When you check a place, we show which events are close enough to reach it.
4 · We make forecasts
Our weather models forecast unusually hot days, unusually heavy rain and a high fire-weather index up to 7 days ahead, at 108 cities, before anything has happened (see Weather models below). Every other forecast starts from an official report: if a hazard such as a flood, storm or fire is still going on, we keep the agency’s alert open for as long as that hazard usually lasts, and say so. After an earthquake of magnitude 5.5 or more, we forecast its aftershocks for 7 days. Every forecast names the hazard, the place and the period it covers.
5 · We lock them
We save each forecast in a public record when we first make it. We only ever add to that record. So the place, the dates and the test for success cannot be changed once the outcome is known.
6 · We check them
When the forecast period ends, we check it. A weather model’s forecast is checked against the weather NASA POWER measured, because a weather line is not a disaster. Every other forecast is checked against the disasters that agencies reported. We mark it correct, partly correct or wrong only after looking at that evidence.
A forecast that restates an agency's alert counts as correct only if new damage or impact is reported inside its period. The agency still listing the hazard is the alert we restated, so it does not count. If our sources could have shown a new impact and show none, the forecast is wrong. These forecasts are not a test of our skill, and we report them apart from forecasts made in advance. Checker version verify-1.4.0.
How we mark a forecast
| Result | What it means |
|---|---|
| Correct | The hazard happened in the place and period the forecast named. For a weather model’s forecast: the line was crossed on at least one day of the window. |
| Partly correct | The hazard happened in that period, but only in part of the place. |
| Wrong | Agencies reported no such hazard in that place and period. For a weather model’s forecast (a false alarm): the line was not crossed on any day. |
| Missed | A disaster happened that none of our forecasts named. For the weather models: a day the line was crossed at a place a model run was watching, with no forecast for it. |
| Cannot be checked | No source we read covers that hazard in that place well enough to say. |
| Test case | The forecast was written after the event had started, so it does not count as made in advance. |
Each hazard has a fixed distance limit, for example 250 km for an earthquake and 300 km for a flood. The same limit applies to every forecast. We do not change it to fit a result. A weather model’s forecast has no distance limit: it is checked at one grid cell.
Each forecast is counted once, by the first version we recorded. A forecast we updated is recorded again, and the later versions are kept but never counted as more forecasts.
Weather models
Our only forecasts made before anything was known. They forecast weather lines, not disasters: a hit means a day crossed a local weather line, not that anyone was harmed. The code is forecast-models.js and tools/run-models.js.
| Inputs | Every day, the Open-Meteo forecast for the next 9 days (daily maximum temperature, rain, mean dew point and mean wind), and NASA POWER’s measurements of the last 60 days, for 108 cities. Each city’s normal is 30 years of NASA POWER records, 1996 to 2025, at the grid cell (about 50 to 70 km across) that contains it. |
| Bias correction | The forecast service and NASA POWER do not agree exactly. Over the last 21 days both cover (at least 15 are needed), the average difference in temperature and dew point is removed from the forecast, and rain and wind are scaled by the ratio of the two (between 0.5 and 2). Where the temperature difference is more than 3 °C, the city is skipped that day rather than corrected. |
| Warming adjustment | The heat line is moved by each city’s own warming trend over 1996 to 2025 (the median slope of its yearly mean temperature, Theil-Sen), to the year it is used in, by at most 2.5 °C. Without it the 30-year line was crossed on 14% of days rather than about 5%. |
| Unusually hot days | The corrected daily high reaches the city’s 95th percentile for the date (a level it reaches about 1 day in 20) on 2 or more days in a row. High if any day reaches the 99th. |
| Unusually heavy rain | Rain in a day at or above the city’s 99th percentile for the month and at least 10 mm, or a 3-day total at or above its 99th percentile and at least 20 mm. High if the topsoil was already wetter than 80% of the same month. |
| High fire-weather index | The Canadian Fire Weather Index, built from the last 60 days of measured weather and the corrected forecast, at or above 21.3 (“very high”) and the city’s 95th percentile for the month, on 2 or more days in a row. High at 38 or more. Fire also needs fuel and a spark, which no weather index can see. |
| Days forecast | Days 1 to 7 after the run. Each forecast is recorded in our public ledger when it is published, before its first day, and never changed. |
| How each is graded | Against NASA POWER for the same grid cell, by the same rule as the forecast: correct if the line was crossed on at least one day of the window, a false alarm if not. NASA POWER revises its recent days, so a day is used only once its measurement has settled: fetched at least 14 days after the day. Until then the forecast waits. Grades and misses are both recomputed from the settled measurements on every run (tools/rebuild-model-grades.js rebuilds them from the published files). |
| Misses and chance | A miss is a day the line was crossed at a city a run was watching (1 to 7 days ahead, and checked that day), with no forecast for it. Every result is also given day by day, as the backtest is scored, and re-issues of one forecast are merged into one claim. Beside each forecast is what chance would give: a window of N days catches a crossing with probability 1 - (1 - p)^N, where p is the share of watched days the line was crossed. |
| Backtest | The method was tested on forecasts Open-Meteo really issued 3 and 5 days ahead, in 108 cities from 17 Jul to 17 Oct 2026, scored day by day against NASA POWER. One season only. Fire was tested with only the last day forecast, an easier test than the live one. Days 1, 2, 4, 6 and 7 ahead were not tested. The results are on the Track record page. |
Types of information
Something that has happened, reported or measured by an agency
Issued by a government or scientific agency
From our weather models, before anything happened, checked afterwards against measured weather
An agency’s current alert, kept open while the hazard usually lasts, checked against new damage reported. Not a test of our skill
Used for background and comparison