FishCast

FishCast models where Antarctic krill (Euphausia superba) are found and how that may change with the climate, especially relative to the marine protected areas (MPAs) managed by CCAMLR. It is a Big Data Big Impact project at Georgia Tech.

The main question for Fall 2026: does the model predict krill better than simple baselines on years it has not seen? See the semester plan for context, and docs/CHANGES_preprocessing_v2.md for the September fixes.

Repository layout

Folder	What's in it
fishcast/	The Python package: data loading, models (GPs, sequence models), training, plotting helpers
scripts/	Command-line entry points: train, benchmark, forecast, visualize
scripts/data/	Data pipeline: download CMIP6, regrid, build krillcast_merged.csv, convert MPA boundaries
jobs/	Slurm job scripts for the PACE cluster
tests/	Automated tests, run on every pull request
data/	Not in git. See data/README.md for where to get each file
docs/	Methodology, model variants, change logs
analysis/	Analysis subteam notes, experiment write-ups, methods research
viz/	Data Visualization subteam: prototype/ polar map, figures/ chart scripts, web app (to come)
notebooks/	01_data_tour and 02_benchmark_results for new members; the old merge notebook for reference

Quick start

git clone https://github.com/gt-big-data/fish-cast.git
cd fish-cast
python -m venv .venv
source .venv/bin/activate            # Windows: .venv\Scripts\activate
pip install -r requirements-dev.txt
pip install -e .
pytest                               # should pass without any data

Then download the data (see data/README.md) and run:

# Train one model (about 5 minutes on a laptop CPU with --epochs 10)
python scripts/train_fishcast.py --model-type latent_state_augmented_gp \
    --data data/krillcast_merged.csv --output-dir artifacts/my_first_run --epochs 10

# Compare trained runs against the baselines on the 2004-2016 test years
python scripts/benchmark_skill.py --gp-runs artifacts/my_first_run --output artifacts/benchmark

How we work

• New here? Start with ONBOARDING.md.
• Before opening a pull request, read CONTRIBUTING.md.
• Tasks live in GitHub Issues on the FishCast Fall 2026 project board. See TASKS.md.

Key results so far

Only models that use recent krill catches currently beat a constant prediction on the test years (about 5% lower error). No climate-only model shows skill yet. All runs from before 27 September 2026 are invalid; see docs/CHANGES_preprocessing_v2.md.