Cause of severe horse feature-union instability

Determine the precise cause of the severe test-performance failure produced by the feature-union procedure for racehorse workout-speed prediction, including the mechanism underlying its test $R^2=-49.112$ on one group-safe split.

Background

The paper evaluates frozen pretrained image encoders and procedures for combining their representations in several prediction tasks, including racehorse workout-speed prediction from standardized conformation photographs. Across twelve group-safe data splits, DINOv2 is selected as the single encoder in every split, whereas feature union is preferred by outer validation in only three splits.

One selected feature-union fit performs catastrophically on the test data, producing a test $R^2$ of $-49.112$, while another produces $-0.099$. The authors note that their training-based numerical safeguard does not constrain extreme test predictions and explicitly leave the precise explanation for this instability unresolved. Identifying its cause could clarify whether the failure arises from feature-scale behavior, distribution shift, a particular encoder block, or another aspect of the feature-union pipeline.