Where weather data comes from and what you may do with it

Every page in this catalogue about a market-data API is about one half of the input: what the venues are pricing. Weather contracts have a second half, and it is a completely different industry with its own products, its own latencies and its own licences.

The confusion that costs people the most is treating "weather data" as one thing. It is at least four, they are produced by different organisations for different reasons, and the one a contract settles on is rarely the one an API sells you by default. What follows is the layer cake, what each layer costs, and which clause in which licence stops you publishing the result.

How it works

Layer one: observations. An instrument at a site reports, and a national meteorological service turns that into products. In the United States that means the Automated Surface Observing System, hourly and special METAR reports for aviation, and the daily and monthly climate text products specified in NWS Instruction 10-1004. These are the products a temperature contract names, and what a temperature market settles on is about reading them as a contract does.

Live access is the National Weather Service API. Its documentation states that "All of the information presented via the API is intended to be open data, free to use for any purpose", requires a User-Agent string identifying your application, and declines to publish its rate limit — it "allows a generous amount for typical use", and an exceeded limit clears typically within five seconds.

Layer two: the station archive. Two NCEI datasets carry the history, and they are not interchangeable.

• ISD, the Integrated Surface Database: hourly and synoptic observations from more than 35,000 stations worldwide back to 1901, of which more than 14,000 are active and updated daily, normalised into one ASCII format. About 600 GB uncompressed. This is the one to use when you need sub-daily values — for example, the hours inside a shifted climatological day.
• GHCNd, the Global Historical Climatology Network daily: more than 100,000 stations in 180 countries, carrying daily maximum and minimum temperature, precipitation, snowfall and snow depth, with about half the stations reporting precipitation only. It takes daily updates and is reconstructed each weekend from more than 25 source components. Its quality control flags about 0.3% of nearly two billion values, and NCEI estimates 98–99% of those flags are true errors.

The number that matters for anyone checking a settlement: in GHCNd, "real-time data sources are typically replaced by archive-ready data sources 45 to 60 days after the end of a month". Before that you are reading the same preliminary value the contract used.

Layer three: raw model output. Numerical weather prediction runs are published as gridded files, free, by both major centres.

• NOAA publishes GFS and the rest of its production suite through its open-data programme, mirrored into public cloud buckets. GFS runs four times a day at 00, 06, 12 and 18 UTC, at a base grid spacing of about 28 km out to one week, coarsening to about 70 km for days eight to sixteen. The registry entry states that "NOAA data disseminated through NODD are open to the public and can be used as desired."
• ECMWF publishes real-time forecasts from both its physics-based IFS and its machine-learned AIFS at 0.25 degrees in GRIB2, under CC-BY-4.0: the data "may be redistributed and used commercially, subject to appropriate attribution". The catch is retention — the open-data archive is rolling, holding roughly the most recent twelve runs, or two to three days. If you want a forecast history you have to capture it yourself, from the day you start.

Layer four: post-processed guidance, and this is the layer people skip. Raw grid output is not what a forecaster uses. In the United States the National Blend of Models combines model inputs and applies statistical correction — decaying-average and quantile-mapping bias correction, with weighted blending — and publishes gridded guidance at 2.5 km over CONUS, updated hourly, with coarser grids for Alaska, Hawaii, Puerto Rico, Guam, oceanic and global domains. Version 5.0 became operational on 5 May 2026. It is free, it is already calibrated, and it is the baseline any home-grown pipeline has to beat rather than the thing a home-grown pipeline replaces.

Sitting beside all four: reanalysis. ERA5 is a model run over the past with observations assimilated into it, at 0.25 degrees and hourly resolution across roughly eight decades, under a CC-BY licence. It updates daily with a latency of about five days, and its early release, ERA5T, can differ from the final version published two to three months later. It is the standard training and verification set for weather models. It is not a record of what a station reported, and using it as one is the most common way to build a calibration that quietly disagrees with the settlement source.

And above them, aggregators. Open-Meteo repackages several of the public sources behind one API, under CC BY 4.0, with a free tier bounded at fewer than 10,000 calls a day, 5,000 an hour and 600 a minute — and that free tier is explicitly non-commercial, with private sites, nonprofit apps, public research and educational content given as the permitted examples. Its paid tiers carry the commercial licence and are sized by call volume: 1M, 5M and above 50M calls a month.

What it costs

The free path is genuinely free, and its real cost is storage and code rather than money. A single GFS run is gigabytes; capturing a rolling forecast archive, decoding GRIB2 and keeping per-station series is an engineering job, not a subscription.

The paid path splits by what you are buying, and the split is the useful thing to notice.

Forecast APIs are cheap and priced per call. Visual Crossing publishes its allowances by record: 30,000 records a month and 1,000 per query on Free, 1,000 records a day on Professional, 10,000,000 a month on Metered, unlimited on Corporate and Enterprise, where "A single 15-day forecast query counts as a single Query Result while each history record returned counts as a Query Result". Open-Meteo's commercial tiers are per-call. Both are self-serve.

Observation feeds sold as settlement-grade are not. The Weather Company — named in August 2026 as the source Kalshi uses to verify weather outcomes — publishes a Standard package at $500 USD per month for 1 million API calls, billed annually, covering current conditions, forecasts, lifestyle indices, alerts and imagery. Its 30-day free trial at 50,000 calls a day is described as "Limited to enterprise customers only". And the two products a settlement check would actually need — "Cleaned historical" and "History on demand" — are listed as enterprise add-ons arranged with its sales team, with no published price.

That asymmetry is the whole economics of this layer, and it runs the opposite way to intuition. Predicting the weather is a commodity: four centres and a dozen vendors will sell you a forecast for the price of a streaming subscription. Establishing, cleaning and licensing a redistributable record of what one instrument reported is the scarce product, and it is the one a contract is written against.

Licences are the part that bites

Nothing above matters until you publish something, and then it matters immediately, because the same chart can be fine under one source and a breach under another.

• US government data is public domain in effect. NOAA asks for attribution on unaltered data and asks that modified data not be presented as NOAA's. There is no redistribution clause to fall foul of.
• CC-BY sources — ECMWF open data, ERA5, Open-Meteo — allow commercial use and redistribution, with conditions. Attribution, and an indication of what you changed. A derived series that drops the attribution is the ordinary way this gets broken.
• Commercial vendors gate redistribution by tier, not by intent. Visual Crossing permits data "storable for shared internal use" on all tiers but "storable for shared external use" only on Enterprise, and requires the line "Weather Data Provided by Visual Crossing" on its Free and Professional tiers. A public dashboard is external sharing whether or not anyone pays for it.

The shape is the one the exchange-data business has had for decades: the numbers are cheap to consume and expensive to pass on, and the licence, not the API key, is what you are actually buying.

What you can do about it

Separate the three questions before you choose a provider. What the contract settles on, what you forecast with, and what you calibrate against are three different datasets. One vendor rarely serves all three well, and picking a single API because it is convenient is how a pipeline ends up calibrated against a grid cell and settled against a station.

For the settlement number, use the exact product the contract names. Not a "current conditions" endpoint, not a nearest-station lookup, not a media summary — the named product from the named source. Kalshi's certified temperature terms say so in as many words: rounding by media outlets, secondary reporting or third-party summaries does not affect resolution.

Build the station history from ISD or GHCNd, not from reanalysis. If you are checking how often a preliminary daily value later moved at the station you trade, you need that station's own record, and both archives are free. Use ISD when you need the hours rather than the daily summary.

Start from the free, already-corrected baseline. NBM is post-processed, station-relevant and published hourly. A pipeline that cannot beat it has not earned its storage bill, and knowing the gap is cheaper to measure than to assume. What an AI weather model gives you is about why raw grid output is not the comparison people think it is.

Read the redistribution clause before you publish, not after. The question is not "may I use this" but "may I show a derived version of this to somebody who is not paying for it". Free tiers answer that differently from paid ones at the same vendor.

Capture forecasts from the day you start. The open real-time archives are rolling — ECMWF's is about two to three days — so a forecast history is something you accumulate, not something you download. If you intend to score yourself against past runs, the cheapest possible day to begin storing them is today.

For the venue-side half of the input — the prices, the order books and the contract metadata — the market data APIs section carries the per-product detail, each card dated to the documents it was read from.