Data

One dataset, with daily and yearly weather for 210 U.S. cities up to 2018-12-31:

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

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()

Metric codes and their columns in annual_indices: annual_avg_temp = Avg Temp, total_precip_in = ToT Prcp, hottest_day_temp = Max Tmax, cdd = cdd.

The dataset page describes each subset and its columns.

How to submit

Your submission is a Hugging Face dataset. Build it like any dataset on the Hub, then submit its link.

1. Create your dataset by copying the template CMU-OSPO/protostar-climate-submission-template into your Hugging Face account.
2. Add your forecasts in the template's submission.csv: the city, year and metric of all 5,880 rows are already filled in; fill in predicted_value (format below). code/check_submission.py in the template runs the same checks as the Space.
3. Write your README.md: your method in a few lines, your Policy Idea and your Feedback to the Ecosystem (see the About tab). It is your write-up.
4. Add your code in a code/ folder (datasets are git repos, so code uploads like any other file).
5. Make the dataset public and paste its link, e.g. https://huggingface.co/datasets/your-name/your-dataset, in the Submission tab.

When you submit, the Space checks that the dataset is public, contains submission.csv and README.md, and that submission.csv has a number for every one of the 5,880 rows; it then records its current version (commit). You can update your dataset and submit again until the deadline; the version recorded at your latest submission counts, so submit again after every change you want judged.

Forecast file format

submission.csv, 4 columns, one row per city × year × metric:

city_id,year,metric,predicted_value
USW00094823,2019,annual_avg_temp,52.4
USW00094823,2019,total_precip_in,38.2
USW00094823,2019,hottest_day_temp,91.0
USW00094823,2019,cdd,845.0
USW00094823,2020,annual_avg_temp,53.1
...
USW00024233,2025,cdd,310.0

• city_id: the station ID from city_info (e.g., Pittsburgh = USW00094823, Seattle = USW00024233).
• Total rows: 210 cities × 7 years × 4 metrics = 5,880 rows.

Starter code

A working model in under 30 lines (scikit-learn):

import pandas as pd
import numpy as np
from datasets import load_dataset
from sklearn.ensemble import RandomForestRegressor

# 1. Load the yearly data (up to 2018)
ds = load_dataset("CMU-OSPO/protostar-climate-train-to-2018", "annual_indices")
train_df = ds["train"].to_pandas()

# Metric codes and their columns in annual_indices:
# annual_avg_temp = "Avg Temp", total_precip_in = "ToT Prcp",
# hottest_day_temp = "Max Tmax", cdd = "cdd"
target = "Avg Temp"

# 2. Build lag features: each city's value 1, 2 and 3 rows (years) earlier
train_df = train_df.sort_values(["ID", "Year"])
for k in (1, 2, 3):
    train_df[f"lag_{k}"] = train_df.groupby("ID")[target].shift(k)
features = ["lag_1", "lag_2", "lag_3"]
fit_df = train_df.dropna(subset=features + [target])

model = RandomForestRegressor(n_estimators=100, random_state=42)
model.fit(fit_df[features], fit_df[target])

# 3. Predict 2019-2025 iteratively
# ... generate predictions and save to submission.csv

# 4. Check your error locally against validation years:
def compute_mae(actual, predicted):
    return np.mean(np.abs(actual - predicted))

FAQ

Can I use data other than the challenge dataset? Yes, with one line drawn by use. Your forecasts may only use records up to 2018-12-31. After 2018, anything that responds to the weather gives away the answers, even when it is not weather data. The same data is welcome in your analysis and policy idea, where real impacts are exactly what a policy targets.

When in doubt: if a 2019–2025 series would be different in a hotter or wetter year, keep it out of your forecasts. List every external source in your README.

Credits

Please cite the data you use (and list any external data in your README):

• Lai, Yuchuan; Dzombak, David (2019). Compiled historical daily temperature and precipitation data for selected 210 U.S. cities. Carnegie Mellon University. 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. 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

My team

Submit

You must be signed in and in a team. Your submission is the link to your team's public Hugging Face dataset, containing submission.csv (your forecasts) and README.md (your method, policy idea and feedback to the ecosystem) and your code in code/. See the Data & submission tab.

When you submit, the current version (commit) of your dataset is recorded. You can submit again until the deadline. Only your team's latest submission counts, at the version recorded then.

Your team's submissions

Judging