Daily precipitation station data and code for "Coupling Precipitation Forecasting and Early Warning with Reverse-Martingale Recurrent Neural Networks"

This dataset contains the daily precipitation records and the complete analysis code that reproduce every table and figure in the accompanying Journal of Hydrology article. The observational data are tidy daily-precipitation series for four contrasting climate regions, each provided as a long-format CSV (date, station_id, precip_mm) together with a station_meta.csv of station coordinates: Taiwan (monsoonal, orographic): 23 stations, 1998-2020, from the public Taiwan Historical Meteorological Observations reconstruction (Raingel/historical_weather) of Central Weather Administration CODiS records; an additional 22-station 2012-2022 subset supports the drought-warning study. Texas Hill Country (semi-arid subtropical): 47 stations, 1980-2024, NOAA GHCN-Daily. Germany (temperate maritime): 50 stations, 2005-2022, NOAA GHCN-Daily. Turkey / Anatolia (Mediterranean-continental): 44 stations, 2008-2019, NOAA GHCN-Daily. The code folder provides the Python implementation of the reverse-martingale recurrent network (RMRNN), the sequential change-point (CUSUM/SR) drought-warning detector, the data retrieval and quality-control scripts, and the domain-calibrated stochastic simulators used for the controlled synthetic experiments. The synthetic-testbed outputs are clearly labelled as such; the headline evaluation is on the real four-region station data above. The results folder holds the produced JSON/NPZ outputs, and README.md documents the run order and exact settings (station lists, year ranges, quality-control thresholds, and per-region date windows). Underlying sources and licensing: GHCN-Daily is public domain (NOAA/NCEI); the Taiwan series derive from the openly available Raingel reconstruction of public CODiS records. The tidy derivatives redistributed here are released under CC BY 4.0.