SAVOS: A Shape-Aware Veto-Based Origin Selector for Operational Forecasting in Monitored Processes

Selecting which forecasting candidate should support an operator is a model-serving decision, yet it is commonly reduced to aggregate error on a single chronological hold-out. This rule has two limitations relevant to operational use. An apparent advantage may not persist across forecast origins, and point-error criteria can favour variance shrinkage on weakly predictable signals, producing low-error forecasts that preserve little of the target dynamics. This study introduces the Shape-Aware Veto-Based Origin Selector (SAVOS), a model-serving selection protocol. SAVOS evaluates candidates across rolling forecast origins, combines predictive skill with shape evidence, and applies eligibility gates before ranking the surviving candidates. If none is eligible, no model is served. Applied to 14 sensor variables from a wastewater treatment plant over 66,590 forecast origins, SAVOS changes 38 of 42 decisions relative to an aggregate-error rule and serves no model in 24. On nitrous oxide, essentially constant forecasts can be significantly more accurate than persistence after multiplicity correction, showing that predictive accuracy and trajectory usefulness are distinct. The results support separating serving eligibility from predictive ranking in monitored-process forecasting.