Skip to content

Scores

How well each model verifies lives on client.scores: the Shyft Score (a composite of many verification lanes, against climatology), its monthly history, regional scores, the monthly ranking across models, badges, and the full per-slice scorecard.

from stormlore import Client

client = Client()
score = client.scores.get("gfs")
score.value                       # e.g. 87.3; None when no score is stored
score.shyft_score.provisional     # True while the window is still filling
score.to_dataframe()              # one row (zero rows when no score is stored)

A model with no stored score is not an error: shyft_score (and value) is None. An unknown model raises ModelNotFoundError.

History and regions

history = client.scores.history("gfs")
history.to_dataframe()            # indexed by month (UTC, month start): value, ci_lower, ci_upper …
history.unreadable                # stored months the API could not read and left out

client.scores.regions("gfs").to_dataframe()   # the nine NCEI climate regions

One month, every model

ranking = client.scores.ranking("2026-10")
ranking.to_dataframe()            # rank (1-based, the API's order), model, value …
ranking.incomplete, ranking.truncated

A month is "YYYY-MM" or a date/datetime (its calendar month; an aware time is converted to UTC first, so 23:30 on 31 October at UTC−5 is November). Anything else raises ValueError before a request is sent.

Badges

badges = client.scores.badges("hrrr")
badges.headline                   # the best badge's id, e.g. "season:summer"; None when none
badges.to_dataframe()             # one row per badge; "rivals" is how many were compared
badges.badges[0].rivals           # each rival with its error ratio

client.scores.badge_headlines().to_dataframe()   # one row per model that holds a badge

Scorecard

card = client.scores.scorecard("gfs", parameter="temperature_2m", lead=0, metric="rmse")
card.to_dataframe()               # one row per entry
card.available.lead_times         # what each filter can take
card.truncated                    # True when the API capped the entries (at 5,000)

Every filter is optional: parameter, lead (hours; 0, the analysis hour, is a real filter and is sent), truth_source, region, metric, month (one calendar month; default the API's rolling window), version, limit. When truncated is true, narrow the filters: the SDK does not page a scorecard.

Scores are served as computed: to_dataframe(units=...) other than "native" raises.

Comparing models

Scores computed on different native grids cover different areas. A model's own scorecard says what it was measured on; read it before putting two models side by side.

Draw it

client.scores.history("gfs").plot()   # the monthly score and its band

Every chart, drawn: charts and maps (pip install 'stormlore[viz]').