The need for uncertainty: why probabilistic LSTMs are key to improving flood predictions and enabling learned warning rules

Code and results accompanying the EGUsphere preprint egusphere-2026-469. This repository implements deep-learning rainfall–runoff models over the CAMELS-CH basins, comparing a deterministic LSTM baseline against three probabilistic output heads — a DRN model (called LD-LSTM in the code), a Bernstein Quantile Network (BQN), and a Countable Mixture of Asymmetric Laplacians (CMAL, kappa-parameterized) — each trained across 5 seeds and combined into an ensemble. Built on top of these forecasts is FRiCA, a PPO-based reinforcement-learning agent for flood early-warning decision-making. Included: hy2dl/ — the core model, training, and evaluation library. notebooks/sb2025/ — all notebooks used to produce the manuscript's results: per-model training, ensembling, and the final plotting/tables notebook. data/FRiCA/ — pre-built episode data for the FRiCA agent. results/sb2025/ — the combined prediction files needed to reproduce the manuscript's tables and figures, plus the final figures themselves. Raw per-seed training outputs are omitted to keep this archive's size manageable. environment.yml / requirements.txt for a reproducible Python environment. Reproducing the manuscript's tables and figures (via sb2025_plots_rev1.ipynb) requires no external data. Retraining models from scratch additionally requires the CAMELS-CH dataset (Höge et al., 2023, https://doi.org/10.5194/essd-15-5755-2023), which is not redistributed here. See the included README for environment setup instructions and a full notebook-by-notebook guide.