Meteorological drought is a recurrent hazard in north-western Iran and a major driver of the water crisis of the Lake Urmia basin. This study forecasts the 3- and 6-month Standardized Precipitation Index (SPI-3 and SPI-6) one month ahead using monthly basin-average precipitation from the Iran Meteorological Organization (2003–2023) together with temperature, pressure, wind speed and GLDAS soil moisture. A climatology-completed SPI proxy, obtained by adding the long-term mean precipitation of the target month to the known part of the accumulation window, was introduced as a physically informed baseline, and the learning models were trained to correct its residual. Multilayer perceptron (MLP), extreme gradient boosting (XGBoost), a gated recurrent unit network (GRU) and three hybrid models (MLP tuned by particle swarm optimization, MLP-PSO; XGBoost tuned by the grey wolf optimizer, XGB-GWO, and by Harris hawks optimization, XGB-HHO) were compared with persistence. The gamma distribution was fitted on the training period only to avoid information leakage. In the 2020–2023 test period the proxy alone reduced the RMSE of persistence from 0.70 to 0.46 (SPI-3) and from 0.51 to 0.36 (SPI-6). The best learning models were XGB-HHO for SPI-3 (RMSE = 0.449, NSE = 0.66) and GRU for SPI-6 (RMSE = 0.354, NSE = 0.77), which improved only marginally on the proxy. Metaheuristic tuning strongly reduced the overfitting of the untuned learners (e.g. the SPI-3 RMSE of XGBoost decreased from 0.58 to 0.45), and MLP-PSO gave the highest drought-class accuracy for SPI-3 (81%). The results show that embedding the structure of the index in the model is more important than the choice of learner for short-lead SPI forecasting.