universe: crypto rows: 352, dates: 44, trials K=159
1. 21-session OOS accuracy (expanding calibration)
width : QLIKE 0.3593 | log-MSE 0.975 | rank-corr +0.489 (n=144) trail21 : QLIKE 0.3477 | log-MSE 0.923 | rank-corr +0.459 (n=144) HAR(5/21/63) : QLIKE 0.3475 | log-MSE 0.917 | rank-corr +0.443 (n=144) w+t21 : QLIKE 0.3428 | log-MSE 0.920 | rank-corr +0.501 (n=144) HAR+w : QLIKE 0.3438 | log-MSE 0.921 | rank-corr +0.473 (n=144) blend(w,t21) : QLIKE 0.3409 | log-MSE 0.913 | rank-corr +0.500 (n=144)
2. multi-horizon (rank corr / QLIKE)
fwd 5d width : rank-corr +0.494 QLIKE 1.1897 (n=152) fwd 5d trail21 : rank-corr +0.486 QLIKE 1.3695 (n=152) fwd21d width : rank-corr +0.489 QLIKE 0.3593 (n=144) fwd21d trail21 : rank-corr +0.459 QLIKE 0.3477 (n=144) fwd63d width : rank-corr +0.535 QLIKE 0.2576 (n=136) fwd63d trail21 : rank-corr +0.537 QLIKE 0.2410 (n=136)
3. incremental (pooled OOS: width + trail21)
log-vol ~ log(width) + log(trail21): b_width +0.411 (t=+2.92), b_trail +0.335 (t=+2.51) b_width date-clustered 95% CI [-0.012,+0.807] (P(>0)=0.97)
4. per-date cross-sectional rank corr(predictor, fwd21)
width : mean +0.415 CI[+0.277,+0.543] (n=42, t=+8.0) trail21 : mean +0.555 CI[+0.479,+0.629] (n=42, t=+13.5) HAR-avg : mean +0.604 CI[+0.512,+0.698] (n=42, t=+16.1)
5. 21-session rank corr by prior trail21 tercile
low-vol (n=112): width +0.277 | trail21 +0.147 | HAR +0.221 mid-vol (n=112): width +0.106 | trail21 +0.165 | HAR +0.148 high-vol (n=112): width +0.394 | trail21 +0.062 | HAR +0.105
6. sizing head-to-head (mom top-decile, 21-session fwd)
equal : mean +8.099% CI[+0.401%,+17.615%] maxDD -40.6% (n=40) inv_width : mean +8.099% CI[+0.401%,+17.615%] maxDD -40.6% (n=40) inv_trail21: mean +8.099% CI[+0.401%,+17.615%] maxDD -40.6% (n=40) inv_HAR : mean +8.099% CI[+0.401%,+17.615%] maxDD -40.6% (n=40)