Mitigating estimation error and sensitivity in dynamic mean–variance portfolio selection

Determine principled approaches to mitigate estimation errors in expected returns and covariances and the resulting sensitivity of portfolio weights in continuous-time mean–variance portfolio selection, so as to achieve mean–variance efficiency over a finite investment horizon in dynamically traded markets without relying on fragile plug‑in estimates.

Background

The paper explains that classical plug‑in approaches to mean–variance optimization suffer from severe estimation error—especially for expected returns—and that optimal weights are highly sensitive to these errors. This problem is exacerbated in dynamic settings where errors accumulate over time. The authors motivate reinforcement learning as a model‑free alternative but explicitly note that broadly mitigating these issues and achieving dynamic mean–variance efficiency remains open.