rows 352, dates 44, K=165
A. 21-session OOS accuracy (expanding calibration)
width : QLIKE 0.3593 log-MSE 0.146 rank +0.489 (n=144) trail21 : QLIKE 0.3477 log-MSE 0.151 rank +0.459 (n=144) EWMA94 : QLIKE 0.3383 log-MSE 0.138 rank +0.505 (n=144) EWMA97 : QLIKE 0.3464 log-MSE 0.138 rank +0.484 (n=144) HAR : QLIKE 0.3475 log-MSE 0.149 rank +0.443 (n=144) EWMA94+width : QLIKE 0.3370 log-MSE 0.135 rank +0.530 (n=144) HAR+EWMA94 : QLIKE 0.3427 log-MSE 0.151 rank +0.500 (n=144)
B. anticipate vol increase: P(fwd21 > trail21) AUC
width rise : AUC 0.465 (n=352, pos=200) width spike50 : AUC 0.418 (n=352, pos=84) trail21 rise : AUC 0.330 (n=352, pos=200) trail21 spike50 : AUC 0.300 (n=352, pos=84) EWMA94 rise : AUC 0.392 (n=352, pos=200) EWMA94 spike50 : AUC 0.335 (n=352, pos=84) HAR rise : AUC 0.358 (n=352, pos=200) HAR spike50 : AUC 0.338 (n=352, pos=84)
C. quantile-skew signal vs forward return
skew IC(fret): mean -0.0065 CI[-0.1095,+0.1024] (n=40) p_up IC(fret): mean -0.0032 CI[-0.1351,+0.1178] (n=40)
Note: GARCH skipped (no arch package); EWMA is the adaptive free benchmark. Caveat: backfill may overlap pretraining.