Agency validation

The model against sourced agency ratings — discrimination works, the letter conversion runs optimistic, and we publish both. Full study →

Discrimination (Spearman ρ vs agency ordering, 90% bootstrap CI)

Stratumnρ90% CI
all names with estimates1280.787[0.71, 0.84]
rated only770.819[0.73, 0.88]
SCALE_RESOLVED only360.726[0.53, 0.86]
Model letters vs sourced agency letters — the letter conversion saturates at the top of the scale.
Model letters vs sourced agency letters — the letter conversion saturates at the top of the scale.
Baselines: plain distance-to-default ties the TiC RiskScore on discrimination; both beat leverage alone.
Baselines: plain distance-to-default ties the TiC RiskScore on discrimination; both beat leverage alone.
RiskScore rank vs agency rank; scale-resolved names emphasized.
RiskScore rank vs agency rank; scale-resolved names emphasized.
Letter notch error: median +5 notches optimistic.
Letter notch error: median +5 notches optimistic.