Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control
Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control
Authors: Anne Somalwar, Bruce D. Lee, George J. Pappas, Nikolai Matni Date: 2026-03-24 Paper ID: arxiv:2603.23465
Summary
This paper analyzes the trade-offs between single-step and multi-step forecasting models, particularly concerning compounding error in learning-based control for linear dynamical systems. The authors theoretically and empirically compare three approaches: single-step models, direct multi-step models, and single-step models trained with multi-step losses. They establish that when the underlying system dynamics are perfectly captured by the model class, single-step models yield the minimal asymptotic prediction error. Conversely, when model misspecification arises from partial observability, multi-step predictors offer significant advantages by reducing bias and improving overall accuracy, a benefit shown to hold in closed-loop control applications.
Key Contributions
- Provided the first quantitative analysis of the trade-off between single-step and multi-step forecasting models for linear dynamical systems.
- Demonstrated that well-specified single-step models achieve the lowest asymptotic prediction error.
- Showed that direct multi-step predictors significantly reduce bias and improve accuracy in cases of model misspecification due to partial observability.
- Provided theoretical and empirical evidence that these trade-offs in prediction accuracy persist when models are utilized in closed-loop control systems.
Limitations
The analysis is primarily focused on linear dynamical systems, and the findings may not directly translate to highly non-linear systems without further investigation.
Open Questions & Future Work
- statistical-efficiency-model-horizon-tradeoff
- asymptotic-prediction-error-comparison-well-specified
- irreducible-bias-comparison-misspecified
- control-cost-decay-well-specified
- stabilization-regimes-misspecified-theory
- extension-nonlinear-systems-theory
- lqr-comparison-multi-step-vs-single-step
- analysis-other-control-frameworks
Limitations
The analysis is primarily focused on linear dynamical systems, and the findings may not directly translate to highly non-linear systems without further investigation.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2603.23465
- paper_id
- 2603.23465
- paper_source
- arxiv
- domain
- time-series
- tags
- time-seriesforecastingevaluationreinforcement-learningcontrol
- architectures
-
- datasets
-
- skill
- TimeSeriesSkill
- created_at
- 2026-03-25T21:18:14Z