A0387
Title: Near-optimal inference and forecasting in time series by using large deep neural networks trained on simulations
Authors: Pablo Montero Manso - University of Sydney (Australia) [presenting]
Abstract: Estimating population parameters from temporally dependent data remains a fundamental challenge in statistics and econometrics. Traditional methods like maximum likelihood, even for ARIMA models, rely on asymptotic approximations, which can lead to bias and unreliable model selection in finite samples. The limitations become more pronounced when time series processes are considered, involving nonlinearities, structural breaks, or non-Gaussian innovations. To address these challenges, a novel simulation-based neural inference framework is proposed. By training a neural network on synthetic data generated from the target model class, the aim is to learn to approximate optimal estimators, bypassing the limitations of traditional analytical approaches. This framework offers flexibility to incorporate desirable estimator properties, including the minimization of user-defined risk metrics or unbiasedness. The efficacy of the approach is demonstrated through comprehensive simulation studies. Crucially, results on real-world datasets show that neural network estimators trained on simple linear processes but optimized to minimize forecast error achieve outstanding performance compared to state-of-the-art sophisticated nonlinear models. Therefore, by leveraging improved estimators, analysts can achieve accuracy, maintain the interpretability of first-principles models, and avoid the need to resort to black-box models.
