Although surface electromyography (sEMG) offers the potential to predict locomotion modes in advance accurately, current methods often provide zero anticipation and disregard the delay caused by computation time. Furthermore, they impose strict window length restrictions to ensure stationarity, thus often requiring additional information, such as integrating more modalities to improve performance. To address these, we initially challenged the prevalent notion that window length should be restricted to a short duration, arguing that it overlooks the untapped potential of advanced end-to-end deep learning models. Accordingly, end-to-end convolutional neural networks (CNNs) were trained for continuous prediction of future locomotion modes, with each model corresponding to one of five different window lengths and one of six prediction horizons-the duration into the future for which a predictive model forecasts a locomotion mode. A comprehensive analysis was conducted on the effect of extended windows and varied prediction horizons with a continuous locomotion paradigm featuring eight locomotion modes involving 16 transitions. Experimental results reveal that a CNN trained with a 2000 ms window and a 250 ms prediction horizon excelled in balancing prediction lead time with accuracy, achieving average accuracies of 96.79 % and 97.14 % in both steady and transient states across eight subjects. Furthermore, the actual anticipatory prediction interval achieved was 247.53 ms, derived by subtracting the average computation time of 2.47 ms from the prediction horizon. These pave the avenue for exploring extended windows with end-to-end deep learning in sEMG-based locomotion prediction, heralding significant potential for future advancements in the field.