✓ arXivScience & Systems
Researchers fix Empirical Bayes' hidden weakness with a new ML hybrid estimator
A new mixed estimator blends Empirical Bayes and maximum likelihood, guaranteeing it is never worse than either at second-order risk.
Empirical Bayes estimators can match maximum likelihood at first-order risk, but researchers find they can underperform plain ML at a finer second-order scale when the estimator's kernel is poorly aligned with the true parameter.
A new paper fixes this by proposing a mixed estimator that blends both approaches, with a closed-form formula for the optimal mixing weight that guarantees improvement over both.
Why it mattersPractitioners using Empirical Bayes without checking kernel alignment may be accepting worse estimates than a simpler ML approach would give.
Open this story in InSnip →
