We study the nowcasting of U.S. state-level fiscal variables using machine learning (ML) models and mixed-frequency predictors within a panel framework. Neural networks with continuous and categorical embeddings consistently outperform both linear and nonlinear alternatives, especially when combined with pooled panel structures. These architectures flexibly capture differences across states while benefiting from shared patterns in the panel structure. Forecast gains are especially large for volatile variables like expenditures and deficits. Pooling enhances forecast stability, and ML models are better suited to handle cross-sectional nonlinearities. Results show that predictive improvements are broad-based and that even a few high frequency state indicators contribute substantially to forecast accuracy. Our findings highlight the complementarity between flexible modeling and cross-sectional pooling, making panel neural networks a powerful tool for timely and accurate fiscal monitoring in heterogeneous settings.