Monthly runoff prediction is crucial for water resource allocation and flood prevention. Many existing methods use identical deep learning networks to understand monthly runoff patterns, neglecting the importance of predictor selection. To enhance predictive accuracy and reliability, this study proposes an RFECV–SSA–LSTM forecasting approach. It iteratively eliminates predictors derived from SSA decomposition and PACF using recursive feature elimination and cross-validation (RFECV) to identify the most relevant subset for predicting the target flow. LSTM modeling is then used to forecast flows 1–7 months into the future. Furthermore, the RFECV–SSA framework complements any machine-learning-based runoff prediction method. To demonstrate the method’s reliability and effectiveness, its outputs are compared across three scenarios: direct LSTM, MIR–LSTM, and RFECV–LSTM, using monthly runoff historical data from Yangxian and Hanzhong hydrological stations in the Hanjiang River Basin, China. The results show that the RFECV–LSTM method is more robust and efficient than the direct LSTM and MIR–LSTM counterparts, with the smallest number of outliers for NSE, NRMSE, and PPTS under all forecasting scenarios. The MIR–LSTM approach exhibits the worst performance, indicating that single-metric-based feature selection may eliminate valuable information. The SSA time–frequency decomposition is superior, with NSE values remaining stably around 0.95 under all scenarios. The NSE value of the RFECV–SSA–LSTM method is greater than 0.95 under almost all forecasting scenarios, outperforming other benchmark models. Therefore, the RFECV–SSA–LSTM method is effective for forecasting highly nonlinear runoff series, exhibiting high accuracy and generalization ability.