Pair leaderboard
Composite IC for every pair from the top-60 features (1,770 pairs).
Bonferroni-corrected p-values — only pairs surviving multiple-testing correction
(n_trials=1,770) are statistically real edges.
Top 30 pairs by |t_NW| receive a full backtest through the production engine.
Showing /
How to read. Composite IC = cross-sectional rank correlation of (sign_a × rank(a) + sign_b × rank(b)) vs forward 252d return.
t_NW = Newey-West HAC-corrected t-stat (autocorrelation-aware).
p (Bonf) = 2-sided p multiplied by 1,770 trials — strict family-wise correction.
Sharpe / Calmar / Skew populated for top 30 by t_NW only (full backtest is expensive).