Table 4.

Mean‐Squared Errors (in (%)²) of the NNs Trained With a 3‐Fold Cross‐Validation Split on the Coarse‐Grained and Preprocessed Quasi‐Biennial Oscillation in a Changing Climate Data

			Type
		Cell‐based	Column‐based	Neighborhood‐based
Neural	Cloud volume fraction	32.77 (28.98)	8.14 (8.03)	25.07 (20.46)
networks	Cloud area fraction	87.98 (80.96)	20.07 (19.79)	52.19 (46.61)
Baseline	Constant output model	684.51	431.28	558.28
models	Best linear model	401.47	97.81	297.63
	Random forest	25.90	161.98	54.74
	Sundqvist scheme	474.12	–	–

Note. Due to computational reasons, only 1% of the data (i.e., ≈10⁷ samples) was used to compute the MSE of the Sundqvist scheme. We only show the MSEs of the models with the lowest loss on their respective validation folds. Here, the neighborhood‐based models comprise one model per split, evaluated on all layers. In parentheses we compute the losses after bounding the model output to the [0, 100]% interval. The baseline models are trained and evaluated on coarse‐grained and preprocessed QUBICC cloud volume fraction data.