Simulacrum Framework Optimizes Time-Series Forecasting and Inference

Time-series forecasting and inference often involve complex challenges like finite-sample bias and miscalibration. Traditional statistical methods can be computationally intensive or analytically intractable for intricate structural models and optimality criteria. Researchers introduce "The Simulacrum," a neural network-based framework that employs decision-theoretic pretraining. Analysts define a generative world and a target decision objective, then train a neural network on stratified simulations from this world. This process allows the network to approximate optimal decision rules. The resulting neural estimator provides forecasts, parameter estimates, predictive intervals, and model selection for zero-shot inference on unseen time series. It consistently outperforms traditional baselines like maximum likelihood estimation and AICc, even achieving state-of-the-art forecasting accuracy on major real-world benchmarks, demonstrating its ability to solve previously intractable time series problems.