rows 3950, dates 79, K=164
A. 21-session OOS accuracy (expanding calibration)
width : QLIKE 0.7807 log-MSE 0.140 rank +0.594 (n=2700) trail21 : QLIKE 0.7496 log-MSE 0.135 rank +0.596 (n=2700) EWMA94 : QLIKE 0.7509 log-MSE 0.120 rank +0.624 (n=1800) EWMA97 : QLIKE 0.7338 log-MSE 0.111 rank +0.660 (n=1800) HAR : QLIKE 0.7418 log-MSE 0.122 rank +0.643 (n=2700) EWMA94+width : QLIKE 0.7597 log-MSE 0.115 rank +0.640 (n=1800) HAR+EWMA94 : QLIKE 0.7411 log-MSE 0.109 rank +0.671 (n=1800)
B. anticipate vol increase: P(fwd21 > trail21) AUC
width rise : AUC 0.425 (n=3950, pos=1939) width spike50 : AUC 0.412 (n=3950, pos=648) trail21 rise : AUC 0.285 (n=3950, pos=1939) trail21 spike50 : AUC 0.296 (n=3950, pos=648) EWMA94 rise : AUC 0.302 (n=2700, pos=1223) EWMA94 spike50 : AUC 0.287 (n=2700, pos=386) HAR rise : AUC 0.337 (n=3950, pos=1939) HAR spike50 : AUC 0.333 (n=3950, pos=648)
C. quantile-skew signal vs forward return
skew IC(fret): mean -0.0471 CI[-0.1003,+0.0080] (n=78) p_up IC(fret): mean +0.0068 CI[-0.0265,+0.0393] (n=78)
Note: GARCH skipped (no arch package); EWMA is the adaptive free benchmark. Caveat: backfill may overlap pretraining.