Penalty kicks often decide championships, yet goalkeepers are left to anticipate the kicker’s intent from subtle biomechanical cues unfolding within a narrow time window. This study presents a real-time, multi-modal deep learning framework for predicting the direction of a penalty kick—left, middle, or right—prior to ball contact. The model adopts a dual-branch architecture: MobileNetV2-based CNN extracts spatial features from RGB frames, while 2D keypoints are processed using an LSTM network with attention mechanisms. Pose-derived keypoints are further used to guide visual focus toward task-relevant regions. A distance-based thresholding method segments input sequences immediately before ball contact, providing consistent input across diverse footage. A custom dataset of 755 penalty kick events was curated from actual match videos, with frame-level annotations for object detection, penalty shooter keypoints, and the final ball placement in the goal. The model achieved 89% accuracy on a held-out test set, outperforming visual-only and pose-only baselines by 14–22%. With an inference time of 22 ms, the lightweight and interpretable design makes the model well-suited for goalkeeper training, tactical analysis, and real-time game analytics.