Waste-generation forecasts are useful only if they improve the collection decisions they inform, yet they are usually judged by accuracy alone. We therefore propagate forecasts through an explicit selection-and-routing pipeline in a synthetic benchmark with 30 realizations (three scenarios, 3052 virtual bins, six waste categories, 730 days). Nine forecasters, comprising eight fixed comparator configurations and one separately tuned bidirectional long short-term memory (BiLSTM)–Nyström kernel ridge estimator (Hybrid), are evaluated on common chronological partitions. GRU attains the lowest mean daily root-mean-square error (0.1437 kg/bin/day) and CalendarRidge the lowest advance seven-day error (0.5949 kg/bin); the Hybrid obtains 0.1477 and 0.6878. Daily plastic forecasts are disaggregated to 60 containers with fixed training-period ratios and drive morning stop selection, followed by capacity- and duty-constrained routing. At 120 kg per vehicle, Hybrid-assisted selection reduces unserved overflowing containers from 6.458 to 0.057 per day relative to reactive collection at the cost of 15.4% more travel, and halves overflow relative to a fixed schedule with no distinguishable increase in distance; at 240 kg, it lowers both travel and overflow relative to the schedule. A naive persistence forecast already captures 49% of the overflow reduction attainable with perfect local information, compared with a 53–58% range for the learned forecasters. Saved predictions and route traces permit independent recalculation. The findings are specific to this simulator and do not establish municipal effectiveness.