Data-Driven Prediction with Stochastic Data: Confidence Regions and Minimum Mean-Squared Error Estimates

TL;DR

This paper develops confidence regions and an optimal minimum mean-squared error predictor for stochastic data-driven prediction, providing statistical guarantees and improved accuracy in control applications.

Contribution

It introduces a data-driven approach that offers statistical confidence bounds and an optimal predictor for stochastic systems, enhancing prediction reliability.

Findings

Confidence regions are valid for stochastic predictors.

The minimum mean-squared error predictor reduces prediction errors.

Numerical results confirm improved accuracy with the proposed method.

Abstract

Recently, direct data-driven prediction has found important applications for controlling unknown systems, particularly in predictive control. Such an approach provides exact prediction using behavioral system theory when noise-free data are available. For stochastic data, although approximate predictors exist based on different statistical criteria, they fail to provide statistical guarantees of prediction accuracy. In this paper, confidence regions are provided for these stochastic predictors based on the prediction error distribution. Leveraging this, an optimal predictor which achieves minimum mean-squared prediction error is also proposed to enhance prediction accuracy. These results depend on some true model parameters, but they can also be replaced with an approximate data-driven formulation in practice. Numerical results show that the derived confidence region is valid and smaller prediction errors are observed for the proposed minimum mean-squared error estimate, even with the approximate data-driven formulation.