SST Baselines

Problem

Before comparing any learned model on sea surface temperature (SST), you need to know how hard simple baselines are to beat. This is a study of those baselines, not a model for operational ocean forecasts. The work was done with the permission of my teacher at HEU.

What I did

Built the data pipeline and evaluation for NOAA OISST v2.1 daily SST in two Pacific regions (Kuroshio and subtropical), 1982–2025. ERDDAP was missing 1,194 days per region, which I filled from NOAA PSL with verification checks. I compared six baselines: persistence, fixed smoothed climatology, anomaly persistence, damped persistence, a 10-year daily climatology (OCN) and per-grid-point Ridge regression. The project has 40 tests and CI.

Method

Fixed split: 1982–2015 for training, 2016–2019 for validation, 2020–2025 for test. The test set was run once, after the configuration was frozen and tagged in git.

Results

Methods that pull toward climatology (damped persistence and Ridge) lost most of their validation advantage in the test years. This is consistent with a larger warm shift in the test period, but the cause is not proven.

Limitations

• Only simple baselines. A learned model (LSTM) is the next step and is in progress; it is not part of the results here.
• Two small regions, one dataset.

AI assistance

Code written with Claude Code; methodology and verification are mine.