FORECAST FLAKES ForecastCase StudyMeteogramsExplorerReplayAbout
FORECAST FLAKES
NBM Case Study · Forecast · Range · Verification

Case study

What was forecast, and what fell

One storm, three questions. A snow forecast is usually delivered as a single number, days ahead of the event, and then never scored. These pages take one storm and show all three halves of it: what the National Blend of Models said at every run from four and a half days out, how uncertain it was in the blend's own published numbers, and what actually fell, on the same grid, in the same 24 hours.

How to read this page
  1. Forecast · storm total. Every frame is the same forecast window. Only the model run changes. Scrub from 108 hours out to zero and watch the blend converge — or fail to.
  2. The range. The blend publishes seven percentiles and ten exceedance probabilities on each six-hour snowfall — but only through f054. Past that there is one number and nothing else, which is why the range chapter is a short-lead chapter and says so. The same chapter carries the blend's own snow level and snow-to-liquid ratio, which is where a snowfall map that looks wrong usually turns out to be right: two towns thirty miles apart get different answers because one is above the rain/snow line and the other is below it.
  3. What fell. The NWS National Snowfall Analysis, 24 hours to 12Z, on a 2.5 km grid.
  4. Verification. Forecast minus observed, and whether the truth landed inside the blend's own 10th-to-90th band. A well-calibrated forecast should land inside it eight times in ten; where it does not, the map says which way.

Every map is a georeferenced PNG drawn through a fixed palette that ships in the catalog, so the pointer readout inverts the exact bytes that were drawn and cannot disagree with the picture. Hover anywhere for the band; turn on City values for the numbers over the towns.

—

Where the probabilities come from. The blend publishes seven percentiles and ten exceedance thresholds on every six-hour snowfall — but only out to f054. A storm total is longer than six hours and quantiles do not add, so the percentile maps here are the six-hour percentiles stacked, and the exceedance maps are read off that stack by inverting it. Stacking is exact only if the windows are perfectly rank-correlated; because that is the widest fan any dependence can produce, the range drawn here is the widest defensible one and the blend's own is no wider. The blend's snow level and snow ratio stop even earlier — their percentiles end at f036 — so only the deterministic value of each is drawn, averaged over the peak day and masked to where that run also had at least a tenth of an inch of snow. A snow level over ground where nothing is falling is not a forecast of anything.
What the verification is. The National Weather Service's snowfall analysis — observed snowfall on a 2.5 km grid built from surface reports, radar-derived precipitation and model guidance. It is the best gridded record of what fell, not a gauge reading: a 2.5 km cell averages over terrain, so a valley town and the ridge above it can share a number that suits neither exactly. The blend is on a different grid again (2345×1597 against 2145×1377), so both are warped onto this raster before anything is subtracted.