What is wired in, and what is not
Forecast Flakes is the winter room of the Forecast network — the place snow, ice and freezing rain get their own treatment instead of being one card on a state portal. It covers the United States and Canada, and it is built to answer the four questions people actually have in a winter storm: when do the first flakes fall, how does the event unfold, how much has to be plowed, and how does this compare to a normal winter.
Forecast Flakes dresses as Forecast City does, the way Forecast Points does: cream atlas paper, the guide-sign green wordmark with its white border, the sign-yellow rule, Overpass for the signage and a serif for anything read at rest. Green is the chrome; blue is the data. The blue is the six-inch band of the National Weather Service snowfall scale, so a bar on a chart, a swatch beside a map's name and the legend under the map are literally the same colour, and it never appears on a button. The two map pages run in Forecast Eagle's dashboard shell — one screen, the maps down the left, the region bar over a map that takes every remaining pixel — because a reader who has the room open all afternoon wants every control on screen.
Most of what this site knows about a place is now a static product: one page and one card per place, for 971 cities, 527 ski areas and 40 mountain passes, rebuilt within the hour of every National Blend cycle and never interactive. The blend's snowfall and ice to ten days with its own 10th–90th percentile range, the day-by-day and the hour-by-hour, the ECMWF ensemble two weeks out, density and plow load, snow level, the trend against the last run, and the blend's own track record this winter against the NWS snowfall analysis — with the run stamp and the source on every one. A briefing is a URL and a PNG, which is what can be saved, printed, sent, or sold.
A winter site is only as honest as its gaps. Rather than fill a range with a model that is not actually running, the status board on the front page marks each one:
Mountain Snowfall takes one ECMWF run (the 0.1° open data, about 9 km underneath, twice a day) and puts it on the real terrain. A 9 km model's mountains are smoothed hills, so three things it hands over are wrong up high. Where the water falls is redistributed by PRISM's 1991–2020 normals inside the half-degree footprint the model actually resolves, keeping that footprint's total — the terrain moves water and never adds any. Whether it is snow is decided from the wet-bulb profile at each 2.5 km cell's real height, not the model's. How deep it piles comes from the Kuchera ratio of the column above that height. Below 2.5 km each cell carries a curve of snowfall against elevation, read at every ski area's lowest and highest lift terminal and at ~60 m terrain over fourteen ranges.
What it does not do: there is no correction for ECMWF's own mountain dry bias (that needs reforecasts this site does not have), the snow ratio is Kuchera's formula rather than a trained model, and the terrain ratio is climatological — a storm from an unusual direction is shared out the way the average storm is.
The Snowfall Explorer is the other half of a forecast: NOAA's National Snowfall Analysis, the operational answer to how much actually fell. Every day since October 2008 at 2.5 km, which is 6,500 daily grids and about 6.5 GB — the average winter, each of the eighteen individual ones, month by month, how often it snows at all, and the biggest day and biggest season ever analysed at every point.
It is published on NDFD grid 184, the same grid the NWS forecast itself uses, which is the point: observed and forecast snowfall line up cell for cell, so verification is a subtraction rather than a regridding exercise. Two things to hold in mind when reading it. A 2.5 km cell averages over terrain, so a lake-effect town and the ridge above it share a number that flatters neither — the snowbelts read perhaps a fifth under their airport gauges. And these are the means of 18 observed winters, not a 30-year normal, because the analysis itself only begins in 2008.
The Snow Situation Room is the same engine running on every National Blend of Models cycle as it lands: snowfall and ice to ten days, by climate day and by horizon, the blend's own 10th–90th percentile range and the chance of an inch, six, or a foot read off it, the snow-to-liquid ratio and the weight that ratio puts on the ground in pounds per square foot, the snow level, and what changed since the last run. All of it on the map, and all of it at 1,500 named places — cities, ski areas with their base and summit, and the mountain passes with the road that crosses them — ranked into boards. Every morning it scores the day before against the NWS snowfall analysis, cell for cell, and adds the row to its ledger.
Everything above is one half of a forecast. The case studies are both halves on one page: the National Blend of Models storm-total snowfall at every run from 108 hours out to zero, the blend's own percentile range and exceedance probabilities around it, and the snowfall analysis it was aiming at, on the same raster and in the same 24 hours. Each storm's peak window is one NOHRSC climate day exactly, so the verification is an identical window rather than an approximately aligned one.
Two storms so far, chosen from the analysis rather than from memory and picked to sit at opposite ends of the winter's problem space. 23–27 January is a synoptic storm over flat, densely populated ground — Oklahoma City to Portland in four days, with 67,763 grid cells over six inches on the 26th alone. 17–20 February is four days of orographic snow where elevation is the forecast: Sacramento finished with nothing at all, Grass Valley at 2,400 ft with 13.8 inches, and Truckee at 5,800 ft — seventy miles up the same road — with 62.2. That second case carries two fields the first does not need — the blend's own snow level and its snow-to-liquid ratio — because no snowfall map explains a difference like that and the snow-level map does.
They are also honest about where the product stops. NOAA publishes snowfall percentiles and exceedance thresholds on the blend only out to f054 — two and a quarter days — and a single number after that; the snow level and snow ratio stop even earlier, at f036. So the range chapters are short-lead chapters and the pages say so instead of filling the gap with something invented. The storm-total percentiles are the six-hour ones stacked, which is the widest fan any dependence between the windows could produce; the exceedance maps are read off that same fan by inverting it, and the inversion is checked against the blend's own published six-hour probabilities on ten million grid points.
Each case has meteograms for its own places, in Forecast Points' interface, that draw snow as snow: every place on the track as a stake filled to what fell, the selected run piling up as a drift whose layers are coloured by how hard each six hours fell, a loom of every run by every six hours so the forecast's changes of mind are a pattern, and a play button that runs the storm. And a locked frame viewer with no map engine at all — three stacked images in a fixed frame — for any machine that cannot hand a browser a WebGL context.
The meteograms answer one more question, and it is the one somebody with a budget actually has: when did we know enough? Each decision threshold — an inch, four inches, the six that means plough it and close the schools, a foot — gets a track of every model run, and a flag at the lock: the earliest run after which no later run changed its mind, and the mind it had was the right one. Two forecasts with the same average error are not the same forecast if one was right from four days out and the other only in the last twelve hours. Truckee is the case in miniature: the blend was nineteen inches short on the amount and yet settled all four thresholds from the very first run, 108 hours out, without one change of mind.
Every place also has a printable briefing — the drift and the stake, the loom, the decision horizon and the written verdict on one sheet — and each case a full PDF with the map of what fell, the picket line of every place as a stake, and every place ranked. They are drawn from the same file the web page reads, with the same colours, so the paper and the screen cannot drift apart.
At the foot of the front page is the 2025–26 season replay: November through March observed against normal, month by month and state by state, with each state's December–February mean ranked against every winter back to 1951. Nationally that winter finished 74th of 75 counting from the coldest — the second-warmest in the record.
Forecast Flakes is a keystone of the Forecast network and will branch into the state and metro sites — Forecast Indiana, Ohio, Texas and the rest — alongside weather.city and squall.cloud.
Three sites already work as one on the case study. Forecast Eagle runs the blend live as national maps and links here for the scored version; Forecast Points owns the meteogram interface the snowgram page is built in and links to it; and the map viewer, the 193-sector table, the 15,000-city label list and 38 MB of cartography are Eagle's, aliased in nginx rather than copied, so the two sites cannot drift apart.
forecastflakes.com · © 2026