Raw model output on the native 0.25° grid — no MOS, no bias correction, no post-processing. Scores understate operational forecast quality. Read the fairness statement →

Data

Everything published here is recomputable from the archived station values, the truth table and this repository at a given commit. Nothing is hand-edited. Errors are stored in °C and displayed in °F on the site; the downloads below are in °C.

schema v0.1
methodology v0.2
data through 2026-08-30
built 2026-08-31T03:52:06+00:00
next update 2026-08-31T11:00:00+00:00

Downloads

File	Contents	Rows	Size
`scores_latest.csv`	every published aggregate: station × model × init × lead × variable × method × window, with n, MAE, bias, RMSE, hit rates, skill and bootstrap intervals (values in °C).	47552	10.6 MB
`pairwise_latest.csv.gz`	paired model-vs-model MAE differences on common days, with the bootstrap interval and the significance flag (°C).	160384	2.0 MB
`daily_errors.csv.gz`	the complete per-day record behind every score: one row per station, model, initialization, lead day, method, variable and climatological day, with the forecast, the observation and the signed error (°C). This is the file to download if you want to recompute anything.	177180	2.0 MB
`stations.csv`	station metadata: identifier, name, truth product, fixed standard UTC offset, coordinates and elevation.	23	1.8 kB
`models.csv`	model registry, including the persistence baseline.	12	792 B

The parquet shards behind these CSVs are mirrored as a Hugging Face dataset: https://huggingface.co/datasets/castcheck/temperature-verification. The code that produces them is at https://github.com/zifanzhang/castcheck (MIT).

JSON API (v1)

Static files, CORS-open, cached one hour. See the API page for worked curl and Python examples and openapi.json.

Endpoint	Contents	Rows/files
`scores/latest.json`	every published score (all stations, models, leads, windows)	47552
`scores/leaderboard.json`	the station_id=ALL slice used by the front page	—
`leaderboard/{window}-{init}z-{method}-{variable}.json`	one pre-built file per leaderboard view (32 of them)	—
`scores/{station}/{model}/{lead}.json`	one permanent-link card plus the last 90 days of daily errors	—
`pairwise/latest.json`	model-vs-model paired bootstrap (station_id=ALL)	160384
`stations.json`	station metadata	—
`models.json`	model metadata	—
`status.json`	pipeline completeness report	—
`openapi.json`	OpenAPI 3.1 description of the endpoints above	—

Table schemas

forecast_values

one extracted station value per model run, valid time, variable and interpolation method (DESIGN §3.1).

Column	Type	Unit	Meaning
`model_id`	string	—	stable model identifier from config/models.yaml
`model_version`	string	—	upstream cycle/version string as advertised by the producer
`init_time`	timestamp	UTC	model initialization time
`valid_time`	timestamp	UTC	forecast valid time
`lead_h`	int16	hours	valid_time − init_time
`station_id`	string	—	ICAO identifier, or ALL for the cross-station aggregate
`variable`	string	—	tmax or tmin (daily extreme of the four common samples)
`bucket_h`	int8	hours	accumulation window of a native extreme field, 0 if instantaneous
`method`	string	—	grid-to-station interpolation: bilinear (headline) or nearest
`value_c`	float32	°C	extracted station value, NaN when missing
`missing_reason`	string	—	why a value is absent; empty when present
`source_url`	string	—	exact object or URL the value was read from
`fetched_at`	timestamp	UTC	when the value was fetched

truth_daily

one row per station-day-source with the first-final policy and QC flags (DESIGN §3.2).

Column	Type	Unit	Meaning
`station_id`	string	—	ICAO identifier, or ALL for the cross-station aggregate
`climo_date`	date	LST day	climatological day, midnight to midnight local standard time
`source`	string	—	truth product: CLI, CF6 or OBS
`tmax_f`	int16	°F	daily maximum as reported by the NWS (whole degrees)
`tmin_f`	int16	°F	daily minimum as reported by the NWS (whole degrees)
`tmax_c`	float32	°C	daily maximum converted to °C
`tmin_c`	float32	°C	daily minimum converted to °C
`issuance_time`	timestamp	UTC	issuance time of the truth product
`is_final`	bool	—	first CLI issued after local midnight (the first-final policy)
`revised`	bool	—	a later corrected report exists (never used in scores)
`revised_tmax_f`	int16	°F	latest corrected maximum, published but not scored
`revised_tmin_f`	int16	°F	latest corrected minimum, published but not scored
`qc_flag`	string	—	quality note, empty when clean
`product_id`	string	—	api.weather.gov product id or IEM archive key
`schema_version`	string	—	data-model version (DESIGN §3)
`methodology_version`	string	—	METHODOLOGY version that produced the numbers

daily_forecasts

sampled and native daily extremes per model run, station and climatological day (DESIGN §3.3).

Column	Type	Unit	Meaning
`model_id`	string	—	stable model identifier from config/models.yaml
`model_version`	string	—	upstream cycle/version string as advertised by the producer
`init_time`	timestamp	UTC	model initialization time
`station_id`	string	—	ICAO identifier, or ALL for the cross-station aggregate
`climo_date`	date	LST day	climatological day, midnight to midnight local standard time
`lead_day`	int8	days	target climatological date − UTC date of the initialization
`method`	string	—	grid-to-station interpolation: bilinear (headline) or nearest
`tmax_sampled_c`	float32	°C	max of the four common samples in the climatological day
`tmin_sampled_c`	float32	°C	min of the four common samples in the climatological day
`n_samples`	int8	count	how many of the four common samples were present (0–4)
`tmax_native_c`	float32	°C	daily max from the model's native extreme field (diagnostic)
`tmin_native_c`	float32	°C	daily min from the model's native extreme field (diagnostic)
`missing_reason`	string	—	why a value is absent; empty when present
`schema_version`	string	—	data-model version (DESIGN §3)
`methodology_version`	string	—	METHODOLOGY version that produced the numbers

scores

published aggregates with bootstrap intervals (DESIGN §3.4).

Column	Type	Unit	Meaning
`station_id`	string	—	ICAO identifier, or ALL for the cross-station aggregate
`model_id`	string	—	stable model identifier from config/models.yaml
`init_hour`	int8	UTC hour	model initialization hour, 0 or 12
`lead_day`	int8	days	target climatological date − UTC date of the initialization
`variable`	string	—	tmax or tmin (daily extreme of the four common samples)
`method`	string	—	grid-to-station interpolation: bilinear (headline) or nearest
`window`	string	—	scoring window: 30d, 90d, 365d or all
`n`	int32	days	number of scored climatological days in the window
`n_stations`	float32	stations	mean number of stations behind each day of an ALL row (1 for a single-station row)
`n_flagged`	int32	days	how many of those days carry a QC flag on the observation
`mae`	float32	°C	mean absolute error, forecast − observed
`bias`	float32	°C	mean signed error; positive = model too warm
`rmse`	float32	°C	root mean squared error
`hit1f`	float32	fraction	share of days with \|error\| ≤ 1 °F
`hit2f`	float32	fraction	share of days with \|error\| ≤ 2 °F
`hit3f`	float32	fraction	share of days with \|error\| ≤ 3 °F
`mae_debiased`	float32	°C	MAE after removing the per-station constant bias of the window; the part of the error that is not a fixed offset
`skill_persistence`	float32	fraction	1 − MAE/MAE(persistence); positive is better
`skill_persistence_debiased`	float32	fraction	the same skill score computed on the debiased errors, so a station with a large constant offset is not scored as skill-less
`mae_ci_low`	float32	°C	2.5th percentile, moving-block bootstrap
`mae_ci_high`	float32	°C	97.5th percentile, moving-block bootstrap
`bias_ci_low`	float32	°C	2.5th percentile of the bias bootstrap
`bias_ci_high`	float32	°C	97.5th percentile of the bias bootstrap
`rmse_ci_low`	float32	°C	2.5th percentile of the RMSE bootstrap
`rmse_ci_high`	float32	°C	97.5th percentile of the RMSE bootstrap
`hit1f_ci_low`	float32	fraction	2.5th percentile of the ±1 °F hit-rate bootstrap
`hit1f_ci_high`	float32	fraction	97.5th percentile of the ±1 °F hit-rate bootstrap
`model_version`	string	—	upstream cycle/version string as advertised by the producer
`segment_start`	date	LST day	first day of the current model-version segment; scores cover only this segment
`period_start`	date	LST day	first climatological day contributing to the window
`period_end`	date	LST day	last climatological day contributing to the window
`computed_at`	timestamp	UTC	when this table was computed
`methodology_version`	string	—	METHODOLOGY version that produced the numbers
`schema_version`	string	—	data-model version (DESIGN §3)

pairwise

paired model-vs-model MAE differences (DESIGN §3.5).

Column	Type	Unit	Meaning
`station_id`	string	—	ICAO identifier, or ALL for the cross-station aggregate
`init_hour`	int8	UTC hour	model initialization hour, 0 or 12
`lead_day`	int8	days	target climatological date − UTC date of the initialization
`variable`	string	—	tmax or tmin (daily extreme of the four common samples)
`window`	string	—	scoring window: 30d, 90d, 365d or all
`model_a`	string	—	first model of the pair
`model_b`	string	—	second model of the pair
`n_common`	int32	days	days on which both models of the pair have a forecast
`mae_diff`	float32	°C	MAE(model_a) − MAE(model_b) on their common days
`ci_low`	float32	°C	2.5th percentile of the paired bootstrap difference
`ci_high`	float32	°C	97.5th percentile of the paired bootstrap difference
`significant`	bool	—	true when the 95 % interval of the difference excludes zero
`method`	string	—	grid-to-station interpolation: bilinear (headline) or nearest
`computed_at`	timestamp	UTC	when this table was computed
`methodology_version`	string	—	METHODOLOGY version that produced the numbers
`schema_version`	string	—	data-model version (DESIGN §3)

Licences and attribution

CastCheck derived tables and this site — CC BY 4.0. Please cite the permanent link of the page you used.
CastCheck code — MIT, https://github.com/zifanzhang/castcheck.
ECMWF IFS HRES and AIFS Single open data — CC BY 4.0, © ECMWF. Source: data.ecmwf.int and the AWS Open Data mirror.
NCEP GFS (AWS noaa-gfs-bdp-pds) — U.S. Government work, public domain.
NOAA/CIRA AIWP model output (AWS noaa-oar-mlwp-data) — U.S. Government work, public domain. Cite: Radford, J. T., et al. (2025). Domain-agnostic evaluation of AI weather prediction models. Bulletin of the American Meteorological Society. doi:10.1175/BAMS-D-24-0057.1.
NWS Daily Climate Report (CLI), CF6 and station observations — U.S. Government work, public domain, via api.weather.gov.
Historical NWS text products — Iowa Environmental Mesonet (IEM) AFOS archive, Iowa State University.

Raw model output, no post-processing. These are values interpolated from each model's native grid, without MOS, bias correction, downscaling or any other post-processing, and they are not the products that ECMWF, NOAA, the NWS or any app vendor issue to end users.

Cite the dataset

CastCheck (2026). Station-level verification of raw weather-model 2 m temperature forecasts [data set]. Methodology version 0.2, schema version 0.1, data through 2026-08-30. https://castcheck.zifanzhang.com

For a specific number, cite its permanent link instead — every score card carries a short and a long citation with the access date.

Changelog

Methodology	Date	Change
v0.1	2026-08-30	First public build: 2 m temperature, 23 stations, ECMWF IFS HRES, NCEP GFS and the NOAA/CIRA AIWP models, persistence baseline, moving-block bootstrap intervals.

Data events (a source outage, a corrected observation, a model version change) are recorded in status and in the gaps list of status.json.

Known limitations

Daily extremes are the max/min of four 6-hourly samples, so they under-state the true afternoon peak and over-state the pre-dawn trough. The bias is identical for every model and is therefore fair for comparison, but the absolute errors are not what a user of a post-processed forecast experiences (METHODOLOGY §2.3).
No elevation or lapse-rate correction is applied; stations whose elevation differs sharply from the 0.25° grid cell carry a representativeness error that is charged to the model.
Truth is the first final NWS CLI report. Later corrections are stored but never change a published score, so a corrected observation leaves a permanent, documented discrepancy.
Models enter the record on different dates, so their windows are not identical. Pairwise comparisons are computed on common days only; the leaderboard columns are not.
Groups with fewer than 30 scored days are published but greyed out and excluded from every ranking; early in a model's record most windows are in that state.
The 0.25° AIWP archive is a research product; outages there appear here as gaps, not as bad forecasts (see /status/).