Analysis code for: Comparative evaluation of statistical and deep learning methods for high-frequency radar surface current forecasting in a narrow tropical strait

Python code that reproduces all tables, figures, and reported numbers of the manuscript 'Comparative evaluation of statistical and deep learning methods for high-frequency radar surface current forecasting in a narrow tropical strait' (Gautama & Putra, submitted to ASCMO). A three-phase benchmark of twelve forecasting methods for HF-radar surface currents in the Sunda Strait: persistence, tidal harmonics (UTide), classical time series (moving average, exponential smoothing, ARIMA), a reduced-rank EOF-VAR spatio-temporal statistical model, shallow machine learning (temporal kNN, perceptron, MLP), and spatio-temporal deep learning (CNN, GRU, CNN-GRU; and for six-hour nowcasting ConvLSTM-ED, BiEF, CNN-GRU-MS). Includes full-domain ARIMA with ADF/AIC/Ljung-Box diagnostics, an ARIMA tidal-vs-residual skill decomposition, multi-seed uncertainty, a controlled same-platform multi-seed (single-GPU, identical config) direct-vs-autoregressive and parameter-matched architecture comparison with training- and single-CPU inference-cost benchmarks, seasonal stratification, gap-fill contamination analysis, vector error metrics, and along-/cross-strait wind-projection analysis. The raw BADA HF-radar and BMKG AWS data are access-controlled under Indonesian national regulation (Peraturan BMKG No. 4/2022) and available on reasonable request to the corresponding author; BATNAS bathymetry is openly available, as documented in the README.