This study broadly investigates a new computational approach called Temporal Basis Function Models (TBFMs) to enhance closed-loop neural stimulation therapies. The research was conducted by analyzing 40 optogenetic stimulation sessions in two rhesus macaques, using micro-electrocorticography to record brain activity and predict responses to paired optical pulses. The findings revealed that TBFMs required less than 20 minutes of data collection and under 5 minutes of training, significantly faster than a linear state-space model (90 minutes) and an autoencoder-LSTM model (over 11 hours). In terms of accuracy, TBFMs achieved a mean test set R² of 0.462 for 164ms forecasts, outperforming linear and LSTM models, and showed stronger discrimination for shorter (40ms) predictions with an R² of 0.787. Simulations demonstrated that TBFMs could effectively detect target brain states and modify neural trajectories, with state-dependence evident in 97.4% of channels. Ignoring the initial brain state reduced prediction performance close to zero, underscoring the importance of this consideration. The discussion highlights that simple, low-latency models like TBFMs offer a practical path toward safe, adaptive neural stimulation therapies by addressing challenges of sample efficiency, training time, and inference speed, paving the way for improved clinical implementations in the future.