CMU-OSPO/protostar-climate-train-to-2018

Daily and annual temperature and precipitation for 210 U.S. cities, up to 2018

Training data for the Protostar Challenge: Climate Risk Intelligence (CMU, November 2026). It is the KiltHub (Carnegie Mellon University) records doi:10.1184/R1/7890488.v6 and doi:10.1184/R1/7961012.v6, cut at 2018-12-31, with a yearly cooling-degree-days column added. Please cite the original records.

Authors of the original data: Yuchuan Lai, David Dzombak License: CC0

Daily maximum and minimum temperature and daily precipitation for 210 U.S. cities with long climate records (each starting between 1851 and 1899), plus yearly averages, totals and extreme indices computed from them. Each city's record is the "threaded" combination of the stations that served it over time (ThreadEx), named by the city's current GHCN station ID. Observations come from NOAA's Global Historical Climatology Network – Daily (GHCN-D), via the Applied Climate Information System (ACIS).

Everything after 2018-12-31 has been removed: the challenge asks teams to forecast 2019–2025 without any data that reveals how those years turned out. Weather records up to 2018 from other sources, and other kinds of data, are allowed (see the challenge's data rule).

Quick start

from datasets import load_dataset

REPO = "CMU-OSPO/protostar-climate-train-to-2018"
daily = load_dataset(REPO, "daily", split="train").to_pandas()
annual = load_dataset(REPO, "annual_indices", split="train").to_pandas()
cities = load_dataset(REPO, "city_info", split="train").to_pandas()
coverage = load_dataset(REPO, "coverage_and_stations", split="train").to_pandas()

Subsets

Subset	Rows	Description
daily	10,245,754	One row per city and day, 1851-01-01 to 2018-12-31.
annual_indices	28,052	One row per city and year, 1851–2018: averages, totals, extreme indices, cdd, and counts of missing days.
city_info	461	Stations combined for each city and the period used from each (a city can have several rows).
coverage_and_stations	28,052	One row per city and year, matching annual_indices: how many days have both temperatures (the days cdd uses), and which station the record comes from.

All four subsets share the station ID (e.g., Pittsburgh = USW00094823) and the city Name; there are 210 cities.

daily

Column	Unit	Description
ID		GHCN ID of the city's current station
Name		City name (no spaces, e.g., KeyWest)
Date		Day
tmax, tmin	°F	Daily maximum and minimum temperature (empty if missing)
prcp	inches	Daily precipitation; trace amounts are 0 (empty if missing)

annual_indices

Column	Unit	Description
Name, ID, Lat, Lon		City, station ID and location of the current station
Year		Year
Miss Tmax, Miss Tmin, Miss Prcp	days	Days with no value that year, for each variable
Avg Temp	°F	(mean daily Tmax + mean daily Tmin) / 2, each mean over the days that variable has a value
Avg Tmax, Avg Tmin	°F	Mean daily maximum and minimum temperature, over the days with a value
ToT Prcp	inches	Total precipitation, summed over the days with a value
Max Tmax, Min Tmax, Max Tmin, Min Tmin	°F	Warmest and coldest daily Tmax and Tmin (Max Tmax = hottest day of the year), over the days with a value
Warm Days, Warm Nights	% of days	Days with Tmax (Tmin) above the city's 90th percentile, reference period = start of record to 1987; percentage of all days in the year, so missing days count as not warm
Cold Days, Cold Nights	% of days	Days with Tmax (Tmin) below the city's 10th percentile, same reference period; percentage of all days in the year, as above
Max 1-day P, Max 5-day P	inches	Largest 1-day and consecutive 5-day precipitation
Wet Days	inches	Precipitation from very wet days (above the city's 95th percentile of daily precipitation)
cdd	°F-days	Added: cooling degree days, see below

Cooling degree days (cdd): the sum over the year of max(0, (T_max + T_min) / 2 − 65 °F), using days with both values. It estimates how much air conditioning a city needs; e.g., Pittsburgh averages 811 °F-days a year (1989–2018), Key West 4,933. Years with no complete days have no value (1,206 rows, mostly early records).

city_info

Column	Description
Name, ID, Lat, Lon	City, current station ID and location
Stn.Name	Name of a station combined into the city's record
Stn.stDate, Stn.edDate	Period used from that station, as dates (end dates capped at 2018-12-31)

coverage_and_stations

How each city-year was measured, not the weather itself. Use it to spot years whose totals read low because days are missing, and jumps caused by a change of station rather than by the climate. Built from daily and city_info.

Column	Description
ID, Name, Year	City and year; the same rows and order as annual_indices
Days Both	Days with both Tmax and Tmin, the days cdd uses (0 where cdd is empty). Compare with the length of the year: 365 or 366 means complete.
Station Segment	Which station of the city's record was in use on 1 January: 1 = first station, 2 = second, and so on, in the order of Stn.stDate in city_info (1 if the record starts later that year). Years with the same number come from the same station.
Station Changes	New stations that started during the year (usually 0). The level of the record can jump in this year.

Example (Denver): Station Segment is 2 until 1995, when the record moves to Denver International Airport (Station Changes = 1); from 1996 it is 3. A trend fitted across 1995 mixes the two stations.

Changes from the original files

• Rows dated after 2018-12-31 removed from daily and annual_indices (the source records run to 2023).
• cdd column added to annual_indices, computed from daily.
• city_info station end dates later than 2018-12-31 shown as 2018-12-31.
• city_info station dates stored as dates (in the source, Stn.stDate was text), and two evident typos corrected: Asheville, ASHEVILLE CITY starts 1876-08-01 (source: 1976-08-01, after its own end date 1964-08-31); West Palm Beach, JUPITER-HYPOLUXO ends 1899-12-01 (source: 1999-12-01, while the next station starts 1899-12-02). No daily values are affected.
• coverage_and_stations subset added, computed from daily and city_info.
• thresholds and linear_regression_results left out: the regression results were fitted on data through 2023.
• Otherwise the files are the Parquet tables of the CMU-OSPO copies of the KiltHub records (the 210 per-city files merged into one table per subset).

Uses

• Intended: forecasting yearly city weather for 2019–2025 in the Protostar Climate Challenge; teaching time-series forecasting with long, real station records.
• Out of scope: describing climate after 2018 (the data stop there); neighborhood-level heat (stations are mostly at airports).

Limitations

• Missing days: totals (ToT Prcp, Wet Days, cdd) and percentages (Warm Days, Cold Nights, …) count only days with a value, so years with missing days read low. Check the Miss * columns and, for cdd, Days Both. The challenge counts a year if no more than 10 days are missing for each variable it uses (Tmax and Tmin for the temperature metrics and CDD, precipitation for total precipitation), and its 2019–2025 values are computed the same way, so training and scoring agree. CDD uses only days with both values: in 18 city-years of 1989–2018, more than 10 days lack one of them although Tmax and Tmin each pass the rule.
• Station changes: each city joins several stations over time (see city_info), so a record can shift when the station changes. Station Segment and Station Changes in coverage_and_stations mark the changes: 32 fall in 1989–2018, and the source records list none in 2019–2023. Sioux City's 1996 change follows 100 years without data (1896–1995).
• Precision: most years are stored in whole °F; 2018 values partly come from converted GHCN-D data and have one decimal.
• No humidity: heat indices cannot be computed.

Citation

Lai, Yuchuan; Dzombak, David (2019). Compiled historical daily temperature and precipitation data for selected 210 U.S. cities. Carnegie Mellon University. Dataset. https://doi.org/10.1184/R1/7890488.v6

Lai, Yuchuan; Dzombak, David (2019). Historical changes of annual temperature and precipitation indices at selected 210 U.S. cities. Carnegie Mellon University. Dataset. https://doi.org/10.1184/R1/7961012.v6

Lai, Y., & Dzombak, D. A. (2019). Use of historical data to assess regional climate change. Journal of Climate, 32(14), 4299–4320. https://doi.org/10.1175/JCLI-D-18-0630.1

Variable definitions and methods are in the authors' README on KiltHub: README.txt.