<p>Existing methods for business process simulation mainly support tactical decision-making, estimating the long-term impact of process changes. In contrast, operational decision-making requires short-term forecasts starting from the current process state. A workaround to tackle this use case is to run long-term simulations up to a point where the workload is similar to the current one (warm-up), and measure performance thereon. This approach, however, does not consider the current state of ongoing cases and resources. We study an alternative approach that initializes the simulation from the current state reconstructed from an event log. We characterize the information a simulation engine needs to start from a current state and propose a method to derive this state from an event log. An experimental evaluation shows this approach yields more accurate short-term performance forecasts than warmed-up long-term simulations, especially under fluctuating workloads. Another challenge in short-term simulation is to communicate the reliability of simulation forecasts. Simulation replications quantify stochastic (aleatoric) uncertainty but not prediction error due to model imperfections (epistemic uncertainty). We propose an approach to quantify the epistemic uncertainty of short-term simulation forecasts. From historical short-term simulations started from reconstructed states, we measure prediction errors for a given performance indicator and learn prediction intervals as a function of workload features. We map start states to workload features, group similar states, and derive an empirical error distribution and prediction interval per group. Experiments show that the resulting intervals are generally well calibrated, and that state-based grouping produces tighter intervals than a global error model.</p>