The National Weather Service publishes its forecast as a grid — the National Digital Forecast Database, the same one behind every weather.gov point forecast — and the calibrated multi-model blend that feeds it, the National Blend of Models. Both cover the lower 48 at 2.5 km and update hourly.

Forecast Points reads a single cell of that grid and shows you everything the forecast says about it: 22 elements on one screen, in your local time, with the nights shaded and the active alerts on top. It is the same idea as the NWS's own IDSS Forecast Points page, built with more room and fewer compromises.

It is the third site on the Forecast Network. weather.city covers places, Forecast Eagle covers the models as national maps, and this covers a point through time. They share a design language and a data pipeline on purpose.

Not from GRIB files fetched per request. Both models are read from weatherzarr, a rolling cloud archive that republishes each NWS cycle as chunked Zarr in two different chunkings — one arranged for maps, one arranged for points.

Forecast Points reads the point-shaped one exclusively, and that is the reason the site is fast. Its chunks hold every forecast hour for a small patch of grid, so a whole meteogram — sixty timesteps of temperature — arrives in a single fetch. Read from the map-shaped layout instead and the same chart costs sixty round trips.

That chunk covers about 264 km square, which has a pleasant consequence: the fetch that answers "Dallas" already contains Fort Worth, Denton and Waco. The extractor is built around it — a run is reduced to the distinct chunks its points touch, so a whole metro area costs about what one point does.

Whole cycles are mirrored to this server as they publish, and every point is then read from local disk. What the site serves is static files: the forecasts are extracted once per cycle and written out, so opening a page costs a file read rather than a query, and there is no service behind it that can be slow or fall over.

Nights are shaded

The cheapest thing that makes a forecast trace readable. The overnight minimum, the timing of the afternoon peak, a temperature that fails to fall after sunset — all of that is shape, and shape is only legible against the day boundaries. Sun times are computed for the exact point, not the timezone.

Every element keeps its own clock

Temperature is hourly, then three-hourly past day three. Probability of precipitation is 12-hourly. Daily maximum temperature is one value a day. They are drawn against real time rather than against a sample index, so nothing is interpolated onto a shared axis it does not belong on.

Bars sit over the window they accumulated in

An NDFD precipitation value stamped 18Z is the rain that fell between 12Z and 18Z, so the bar is drawn to the left of its stamp and as wide as its own accumulation window — which is 6-hourly early in the run and 12-hourly later. Centring the bar on its timestamp, the obvious thing to do, puts every rainfall in the country several hours late.

Probabilities are blocks, temperatures are lines

A 12-hour chance of precipitation holds flat across its window; drawing a line between two of them would show a smoothly rising chance of rain that the forecast never asserted. Temperature between two hourly stamps really is approximately a line. Different data, different mark.

Daily highs and lows are dots, not a curve

There is one value per day and it describes an afternoon or an overnight, not the instant it is stamped at. Joining them would draw a temperature trace that no element in the forecast contains.

Axes have floors and minimum widths

A day when visibility never leaves ten miles has a data range of zero, and naive auto-scaling would render that flat line as dramatic noise across an axis spanning a thousandth of a mile. Non-negative quantities sit on zero and every element declares the narrowest axis it may be drawn on.

It stays on paper

The brief was a trading terminal — dense, monospaced, numbers in columns — and the layout takes that seriously. But terminals are dark because they are read in a dark room for eight hours. A forecast is read for ninety seconds in daylight and then printed or pasted into a briefing, so this is a document: light ground, one palette, and a print stylesheet that drops the chrome and lets the chart fill the page.

Two NDFD elements — weather type and hazard — arrive from the archive as bare integers, with no lookup table and no GRIB parameter id to resolve one from. The publisher has already collapsed NDFD's compound weather string into a number whose key we do not have.

Inventing a legend for those would print words like "snow" under a forecast on the strength of a guess, so they are passed through as opaque codes and left out of the chart. The hazard information you actually want comes from the live NWS alerts feed instead, which carries real event names and the official colours.

Snowfall and ice accumulation are published, converted and drawn, but their units are flagged unverified: they were identically zero across every tile sampled when this was built in mid-August, so the millimetre assumption is inherited from the rest of the precipitation family rather than measured. Rainfall was measured and is correct. The flag comes off when a winter cycle settles it.

And nothing is interpolated. NDFD is already a forecast grid — the values are what a forecaster drew — so the nearest cell is reported as-is, along with its own coordinates and how far it sits from the point you asked about. On a 2.5 km grid that is at most about 1.8 km.

The network's shared visual language: warm paper, Overpass — the open metric-compatible cut of FHWA Highway Gothic, the actual Interstate signage face — and MUTCD colours used the way the manual uses them.

Forecast Eagle leads with guide-sign green and marks itself with the Interstate route shield, because it is about the national network of models. Forecast Points leads with motorist-services blue — the colour of the sign that tells you what is available at this exit, rather than where the road goes. That is exactly what a point forecast is: everything on offer at one location. Green stays as the second voice, so the two sites are plainly the same network without being the same page.

There is no dark mode, here or anywhere on the network.

This is not an official source. It is an independent reader of public NOAA/NWS data. For warnings and official forecasts go to weather.gov.

The underlying data is US Government work and in the public domain, and it is published by NOAA/NWS for anyone to use — go to weather.gov or the NWS API for it directly. What is generated here feeds this site and the rest of the Forecast Network; it is not offered as a data service.

Built by Ryan Maue.