Reliable monthly streamflow forecasts support reservoir operation, water allocation and drought preparedness, yet the skill of data-driven models is rarely compared across hydrologically contrasting catchments under an identical protocol. This study evaluated multiple linear regression (MLR), support vector regression (SVR), random forest (RF), gradient boosting machine (GBM) and a multilayer perceptron (MLP) for monthly streamflow prediction in six French catchments (Bruche, Meuse, Loing, Odet, Ubaye and Esteron) that span oceanic, lowland karstic, groundwater-dominated, mid-mountain, Alpine nival and Mediterranean regimes. Daily SAFRAN-based precipitation, temperature and potential evapotranspiration and observed streamflow from the CAMELS-FR sample (1999–2018) were aggregated to 240 monthly values per catchment. Three input scenarios were built from autocorrelation and cross-correlation analyses: antecedent flows only (S1), antecedent flows plus antecedent climate and a seasonal signal (S2, a true one-month-ahead forecast), and S2 plus same-month climate (S3, a perfect-forcing upper bound). Models were tuned by five-fold forward-chaining cross-validation on 1999–2012 and tested on 2013–2018 against persistence and climatology benchmarks. In the forecast setting (S2) the best models reached a Nash–Sutcliffe efficiency (NSE) of 0.25–0.87 and beat climatology in every catchment, but gains over MLR were small. Adding same-month climate (S3) raised the best NSE to 0.63–0.94 and reduced RMSE by 21–65%, the largest gains occurring in the rain-fed Esteron, Odet and Bruche catchments and the smallest in the snowmelt-driven Ubaye. The MLP ranked first in S3 (mean NSE = 0.85; best in five of six catchments), followed by SVR (0.80), whereas RF and GBM overfitted and could not extrapolate beyond the largest training flows. Permutation importance showed that one-month lagged flow and same-month precipitation dominated, while seasonal terms and 12-month lagged flow were informative only in the nival Ubaye. Forecast skill at the monthly scale is therefore governed more by catchment memory and by the quality of meteorological forecasts than by the choice of algorithm