Affected by periodic reservoir water level fluctuations and seasonal rainfall, most reservoir landslides show a step-like pattern of dynamic switching between accelerated and decelerated deformation states. This paper proposes a novel adaptive interval prediction method for accurate prediction of such step-like landslide displacement. In this method, future deformation state (acceleration or deceleration) of landslide is identified by random forest model. Based on the dual-output least squares support vector machine (DO-LSSVM) model, two independent interval predictors are established to adaptively predict landslide displacement in the accelerated deformation state and the decelerated deformation state, respectively. Landslide triggering factors, including rainfall and reservoir water level fluctuations, are adopted as inputs to train the random forest model and DO-LSSVM predictors. To depict the effectiveness of the proposed method, a typical step-like landslide, the Baishuihe landslide in the Three Gorges Reservoir region of China, is taken as a case study. The prediction performance of landslide displacement is evaluated by the prediction interval coverage probability (PICP), the normalized mean prediction interval width (NMPIW), and the modified coverage width-based criterion (CWC). Compared to existing interval prediction methods, the predicted displacement intervals by the proposed method have smaller NMPIW and relatively high PICP. Moreover, the mean and standard deviation values of CWC are much smaller than those obtained from existing methods, showing improved prediction accuracy and reliability. Results of this study confirm good performance of the proposed method in interval predictions of step-like landslide displacement. The prediction results could facilitate early warning of landslide disasters.