"EqMom improvement probes (2026-10-04) - holding-period + top-K sweeps (no code changes)"

#Hold-period sweep (101-name live universe, 12m raw momentum, equal weight, 10bps, next-open)

• hold 7: OOS ret 953.6%, SR 1.02, maxDD -34.0%, turn 10.6x
• hold 21: +1240.8%, SR 1.10, maxDD -35.5%, turn 5.5x (live)
• hold 42: +1356.1%, SR 1.11, maxDD -35.6%, turn 4.0x
• hold 63: +1257.7%, SR 1.08, maxDD -35.4%, turn 3.2x -> 42d marginally best SR, but within noise; hold-21 stands.

#top-K sweep (same universe, same signal, 21d hold)

• top_k=5: OOS ret 2581.6%, SR 1.16, maxDD -37.9%
• top_k=10: +1240.8%, SR 1.10, maxDD -35.5% (live)
• top_k=15: +715.0%, SR 1.00, maxDD -32.4%
• top_k=20: +597.2%, SR 0.99, maxDD -32.2%
• top_k=30: +575.7%, SR 1.04, maxDD -33.5% -> k=5 is SR-better but concentrates (5x names) and MDD deeper; not enough to adopt without a paired bootstrap. Because production supports EQMOM_TOP_DECILE=0.05 env, a paper-trial at 5% is realistic; if it graduates I'd formalize as eqmom_top5 variant.