Predictor Statistical Efficiency Tradeoff

Background: In learning-based control, compounding prediction errors from recursively applied single-step models pose a challenge for long-horizon forecasting and controller design. Direct multi-step predictors mitigate this by learning a single model for the entire horizon, but this increases model complexity and data requirements.

Question / Future Work: Analyze the trade-off between statistical efficiency (sample size needed for convergence) and predictive accuracy when comparing single-step models rolled out autoregressively against directly trained multi-step models, especially under system misspecification due to partial observability.