We ship a close-based structural stop, breakeven at +8% and a 20-day trail. MarketSmith ships a fixed 7–8% loss rule and a 20–25% profit target. This replays 192 cells across six levers on the six-month book's own daily bars — through one ₹60L account with 12 slots, so a rule that holds a position pays for it by missing the next signal.
The first sized every signal independently, with no cash pool and no slot limit. Measured afterwards, the arms needed ₹339L across 68 simultaneous positions against a ₹60L book. Under unlimited capital holding is free, so the arm that never sells always wins — the ranking was ordering arms by how much imaginary money they borrowed, and its top arm is near the bottom here.
The second still marked two names on stale dates. BLISSGVS carried ₹4.49L of MarketSmith's total — 19.5% of that arm — marked at its 2 July close because the bhavcopy has no later bars for it, on a name that in the real book was a same-day scratch for −₹175. BLISSGVS and HFCL are excluded here, which costs MarketSmith 4.6 points and leaves our own arms untouched.
Hold sizing fixed at the book's own quantities and our exit rule is worth +3.3 points over what actually happened. Change the sizing to today's rule — ₹40,000 of risk, ₹5L a name, 12 slots instead of the 35 thin positions the book ran at its April peak — and that becomes +14.8 points. Roughly three and a half times as much of the improvement is size, not exits.
| arm | today's sizing | book's own sizing |
|---|---|---|
| ours (shipped) | 20.9% | 9.4% |
| MarketSmith | 33.9% | 15.0% |
| ACTUAL BOOK, real quantities | — | 6.1% |
| arm | realised | open | total | return | CAGR | max DD | taken | missed | win % |
|---|---|---|---|---|---|---|---|---|---|
| ACTUAL BOOK | — | 3.64 | 6.1% | — | 132 | 37 | |||
| ours (shipped) | 10.29 | 2.25 | 12.55 | 20.9% | 49% | -5.1% | 48 | 79 | 52 |
| ours, no trail | -3.21 | 15.60 | 12.39 | 20.7% | 49% | -10.3% | 23 | 104 | 43 |
| MarketSmith | 14.39 | 5.95 | 20.34 | 33.9% | 85% | -5.0% | 30 | 97 | 67 |
| MS, no target | -2.21 | 15.30 | 13.09 | 21.8% | 52% | -6.9% | 15 | 112 | 67 |
| MS stop + our trigger | -1.73 | 15.49 | 13.77 | 22.9% | 55% | -8.0% | 14 | 113 | 71 |
Equity is marked to market daily, and every curve reconciles to ₹60L plus its own P&L. CAGR annualises a 173-day window and should be read as arithmetic, not as a forecast. Win rates include unsold positions, which cannot have lost yet: MarketSmith's 67% is 30 trades of which 10 are still open.
A lever measured on six months that helps in April–May and hurts in June–September is a fact about those months. Faded bars reverse sign between the halves. What is left standing is stop pct5 vs structural, trigger intraday vs close, target t25 vs none, trail ma20 vs none.
| lever | whole period | Apr–May | Jun–Sep | verdict |
|---|---|---|---|---|
| stop pct5 vs structural | -0.18 | +0.13 | +0.37 | agrees |
| stop pct7 vs structural | -0.40 | -0.32 | +0.77 | sign flip |
| stop pct8 vs structural | -0.30 | -0.07 | +1.10 | sign flip |
| trigger intraday vs close | +2.22 | +0.21 | +1.58 | agrees |
| target t20 vs none | -0.76 | -7.22 | +4.00 | sign flip |
| target t25 vs none | -4.15 | -6.99 | -0.03 | agrees |
| breakeven be8 vs none | -1.29 | -2.11 | +0.97 | sign flip |
| trail ma20 vs none | +4.88 | +1.11 | +0.88 | agrees |
| 8wk hold on vs off | +2.68 | +2.77 | -0.49 | sign flip |
Dropping the 20-day trail returns 20.7% against our 20.9% — almost identical — at -10.3% drawdown against -5.1%. It also takes 23 trades where we take 48, because a position that is never sold never frees its slot. As a paired lever across all 192 cells the trail is worth +4.88pp and it agrees in both halves. This is the lever the unconstrained version of this study scored as negative; recycling capital is invisible when capital is free.
| period | signals | ours | MarketSmith | MS wins, of 200 orderings |
|---|---|---|---|---|
| Apr-May | 87 | 13.3% | 15.6% | 110 / 200 |
| Jun-Sep | 40 | 3.7% | 11.4% | 200 / 200 |
On the book's own queue order MarketSmith makes ₹20.34L against our ₹12.55L — a gap of ₹7.80L. That single number is worth very little on its own: run the same comparison on 200 random tie-break orders and MarketSmith wins 186 of 200, with a median edge of +8.3pp, a tenth percentile of +1.5pp and a worst case of -7.9pp. The gap is real on average and routinely negative.
Split it and the direction reverses outright: ours wins April–May 90 times out of 200 by a median -0.7pp, theirs wins June–September all 200 by a median +8.6pp. Two rules that each win half the sample decisively are not one rule beating another; that is a regime effect, and it is the same verdict the sign-flipping target and 8-week hold already point at.
Corrected 25 Sep, evening. These win counts previously read 200/200 and 34/200. That comparison sorted both sets of runs before comparing them, so it asked whether MarketSmith's k-th best run beat our k-th best run — stochastic dominance between two distributions, not a win on the same queue. Paired properly, the whole-period claim falls from 200/200 to 186/200.
Ours ranks 133 of 192 cells, and the leaderboard is not the evidence — picking the best of 192 is a selection, not a measurement. The main-effect table with the split is the evidence, and it supports changing nothing about our exits. The one durable difference between the two systems is that MarketSmith's stop triggers intraday and ours waits for the close; that lever agrees in both halves and is the only part of their exit set that is both theirs and survives.
Nothing was changed on the strength of this page. The one rule change made on 2026-09-26 was recording the 12 slots the account already ran, which is a sizing fact and sits on the other side of the decomposition above.