Tag
This paper introduces algometrics, a framework for time series forecasting under algorithmic feedback, proving that deployment risk differs from historical risk and is not identifiable from passive data alone. It provides methods for estimating deployment risk using interventions or randomized actions.