Good suggestion, thanks. The per-date half I can answer from the write-up: on the 2025 to July 2026 test, mean per-date Spearman is +0.3452 for raw-feature ridge and +0.4145 for the transformer ensemble, against +0.3428 and +0.4143 pooled. So the gap holds date by date, and pooled versus per-date barely moves for any of the methods here.
Ranked ridge, with the features and the target both ranked within each settlement date, I haven’t run. Two things hint at where it might land. The three-factor composite is already an unweighted average of within-date ranks and gets +0.3057 per date. And in a later direction-accuracy experiment, swapping the window aggregates for within-date rank or z-score views scored slightly below the raw values. My guess is that some of the signal sits in levels a rank throws away, like flow relative to the stock’s own prior short interest, but ranked ridge is the clean way to measure how much a linear model gets once scale is out of the picture. It’s a better challenger than raw ridge, so I’ll treat it as the baseline to beat next time.