Bias of FCNN and RF on the corresponding test set, composed of four theoretical SN Ia models used for training, and a data set based on the empirical SNEMO15 model, not used during training, for a variety of different hyperparameters and filters (i, z, and y), i.e., from model averaging. The large biases on the SNEMO15 data set up to 1 day in our FCNN approach come from the different hyperparameters even though the corresponding test set provides biases within 0.25 days. The RF provides much lower biases in all cases; it depends only weakly on the hyperparameters and is instead mostly set by the filters under consideration.