Code
Open-source code released alongside our publications for full reproducibility.
We release the code behind our studies so that results can be inspected and reproduced. Each entry links to a public repository and the paper it accompanies.
ACES — Adaptive Conformal Early-warning System
MIT LicenseReproduction code for "Calibration, not loss design, governs prediction-interval quality: an adaptive conformal early-warning system for algal-bloom forecasting." The benchmark crosses five heteroscedastic loss families with three forecasting backbones (Mamba, GRU, inverted Transformer), pairs every predictor with raw, split-conformal, and online adaptive-conformal calibration, and evaluates an early-warning decision layer on the calibrated intervals. The central finding: the calibration layer — not the training loss — governs prediction-interval quality under distribution shift.
Uses publicly available observations from the Korean national water-quality monitoring network (water.nier.go.kr), Geum River basin, 2021–2024. Raw records are not redistributed; see the repository README for the preprocessing pipeline.
View on GitHub