Forecasting under Persistent-Transient Ambiguity
Abstract
Online forecasting methods increasingly emphasize rapid adaptation to concept drift, yet newly observed behavior may indicate either the beginning of a persistent distribution shift or a transient episode caused by an unusual event, sensor failure, or malicious poisoning. These alternatives may be indistinguishable at the moment the data is received, although they imply opposite decisions for future deployment. We therefore formulate adaptation as an operational deployment decision problem. A newly updated model is kept only if its out-of-sample forecasts demonstrate statistically credible improvements compared to the previous model. We formalize the cost of this ambiguity by showing that any model promotion rule before observing subsequent outcomes incurs a nonzero worst-case excess risk across indistinguishable persistent and transient continuations. The resulting framework shifts the emphasis from unconditional rapid adaptation toward cautious updating by hedging under ambiguity and delaying decision with statistical evidence.