Table 3.

Top models from each level showing substitutions between cardinal directions and time layers. Each 500K training draw is from a population of 4 million, with an accompanying 500K test set, randomized before split. All three data sets agreed on the top models for levels 1 through 3. Level 4 substitutions shown in red.

Level	500K-A	500K-B	500K-C
6	IV:N1E3Z5:W2Z5:S3Z5:Z4Z5	IV:W1S3Z5:N2Z5:E3Z5:Z4Z5	IV:N1E3Z5:W2Z5:S3Z5:Z4Z5
5	IV:N1Z5:W2Z5:E3Z5:S3Z5:Z4Z5	IV:W1Z5:N2Z5:E3Z5:S3Z5:Z4Z5	IV:N1Z5:E2Z5:W3Z5:S3Z5:Z4Z5
4	IV:W1Z5:S2Z5:N3Z5:Z4Z5	IV:W1Z5:N2Z5:S3Z5:Z4Z5	IV:W1Z5:N2Z5:S3Z5:Z4Z5