Estimating win probability during live basketball games requires integrating both the current game state and the sequential history of play-by-play events. We propose a neural architecture that learns dense embeddings from play-by-play event sequences and combines them with game state features to produce probability estimates. The model encodes each event using concatenated embeddings for event type, team attribution, and temporal dynamics, then applies a recency weighted pooling mechanism that emphasizes recent events while retaining information from the full sequence. A fusion layer combines the sequence representation with a game state embedding that encodes score differential, time remaining, and team strength priors. We evaluate our approach on 500 NBA games against the ESPN Analytics model, a widely deployed industry baseline. Our model improves log loss and Brier score under trajectory based evaluation, and it achieves lower expected calibration error under whole game evaluation across all events. The performance gap widens as games progress, indicating that accumulated event history provides increasingly valuable predictive signal over time. The results support the value of sequential modeling of play-by-play events beyond what instantaneous game state features express.