HydroSPI-ARX for explainable multi-scale drought early warning over Türkiye from ERA5 and ERA5-Land reanalyses

Abstract Machine-learning drought studies typically regress a smooth, autocorrelated index at single stations and report near-perfect fits. Temporal leakage is rarely controlled, and the skill the index already carries in its own history is rarely separated out. Here we present HydroSPI-ARX, an autoregressive and explainable framework for drought early warning over Türkiye. It recasts the Standardized Precipitation Index (SPI) as a leakage-controlled, lead-resolved classification target built from ERA5 and ERA5-Land hydroclimatic fields. Five drought targets are evaluated at leads of one, three, six and twelve months, with feature selection and scaling recomputed inside every training fold. The moderate-or-worse warning reaches a macro-F1 of 0.853 one month ahead, 0.794 at three months and 0.709 at six months. At twelve months it falls to 0.486, where no configuration retains skill. Naive persistence is ahead at one month; against it the framework gains 0.038 macro-F1 at three months and 0.066 at six, with year-block bootstrap intervals that exclude zero. Against a per-cell linear autoregression on the index's own memory, the gain is measurable only at six months (0.043, 95% interval 0.010 to 0.078). Exogenous hydroclimatic fields add 0.007 at one month and nothing beyond. Skill is therefore carried by the multi-scale memory of the index, and the learned model earns its place at the six-month horizon. Scoring the same held-out forecasts against station-derived labels lowers agreement (Cohen's kappa of 0.30, 0.26 and 0.21 at one, three and six months). The ranking of the forecasts survives, so the advantage is not an artefact of verifying against the same reanalysis. ERA5 and ERA5-Land are equivalent within a margin of 0.02 macro-F1 set in advance, and a model trained without a climate zone still forecasts inside it. A Mann-Kendall analysis shows rising three-month drought frequency over 1981 to 2025.