The Filter We Skipped — Qullamaggie’s Strong-Market, Hot-Sector, Market-Wide-Leader Filter, Applied by a Machine

What we granted, and what we actually tested

Granted, not measured: by his own published record and by our first episode’s verdict, that Kristjan Qullamaggie trades this profitably, and that the published breakout mechanics reproduce the profile he describes — a low win rate and a fat tail, the median trade losing about what it risked. None of that is on trial here. His read of which theme is driving a group, his entries during the trading day and his sizing by conviction are not written down as rules, so they are not in the code, and a filter without them is his filter, not his edge.

Tested: one thing: whether the filter the viewer named — as Qullamaggie himself states it on his nightly scans (the stock in the top 7% of the whole market by return over one, three or six months; in an industry group whose three-month return is in the top quarter of groups; in an uptrending market) — recovers an edge for the published entry when applied mechanically. Five arms on one footing: the control, then each rung added, then the sector rung on its own to isolate which half carries anything. 38 trade ledgers, 7,734 trades, all of them below or downloadable, none dropped.

§1 · The locked pre-registration

Committed d755517 on 2026-09-14, before the probe was run — the wager you cannot move after seeing the cards.

The falsifiable question

Does the strong-market + market-wide-leader + hot-sector filter, applied mechanically to the published breakout entry on a point-in-time, survivorship-free universe, recover a positive edge — and does that edge survive both halves of the window, the removal of its three best names, every neighbouring dial setting, and a holdout the filter never saw?

• A yes looks like: the full filter clears PF 1.3 with an average trade of +0.30 R or better on at least 150 trades, holds in both halves and without its top three names, keeps PF above 1.0 at every dial setting, and reads PF 1.1 or better on the 2010–2015 holdout.
• A no looks like: the filter reads at or below the plain entry on the pre-registered rulers, or clears the first gate in sample and dies out of sample — the pattern this channel has watched repeatedly.
• Ambiguous residue: an arm where the two rulers disagree — a positive average trade beside a profit factor under one. That happened on the control and on the sector-only arm, and it is reported as two rulers printed side by side, never as one of them read alone.

Our stated prior: a coin flip on the first gate, fifteen in a hundred on all four

Written beside the pass mark before the run: P(G1) about 0.5 — a leader filter should raise the average trade in sample — and P(G1 through G4) about 0.15, because this channel has watched in-sample lifts die out of sample before and the second half of the window holds a bear market. The likeliest honest story we wrote down was that the filter beats the control in sample, the lift is carried by three names or by the 2020–21 stretch, and the holdout kills it. The result embarrassed the prior in the other direction: the full filter did not beat the control on any registered measure even on the data it was built on.

The primary configuration (locked)

Deliberately NOT tested — named in advance

1. His judgment. Which theme is driving a group, which chart is worth his time, how much conviction a position deserves — none of it is written down as a rule, so none of it is in the code. The verdict is about the rule.
2. A better sector map. The two-digit SIC code is the label stamped on every listed company and it is coarse; a point-in-time theme classification is the first reopen criterion, not something this lab claims to have tried.
3. Intraday entries. The entries are end-of-day buy-stops, as in the first episode; his are placed during the session.
4. The first episode’s own exclusions, carried. The hand-picked name list, the uncapped breakout-day-low stop, and any universe built from today’s survivors are not evidence here; the survivors-only universe appears once, as the control column of the parity table, so the flattery can be seen.

The goalposts — all required to proceed

0 of 5. The first gate asked for three things at once and the full filter missed all three — PF 0.55 against 1.3, an average trade of -0.33 R against +0.30, 78 trades against 150 — and none of the misses was close. The other four gates are printed because the pre-registration said every gate ships whether or not it flatters the concept, and each one fails on its own: both halves under 1.0, the ex-top-3 cut at 0.29, the holdout at 0.54 on 45 trades, and a dial grid whose best setting reads 0.97 in one place and moves to a different setting in the other half.

§2 · The headline arithmetic

The ladder, primary window

The same five arms on the 2010–2015 holdout, run once

Profit factor ($) is gross profit over gross loss in portfolio dollars, the goalpost metric as the first episode defined it; profit factor (R) is the same on R-multiples, size-independent. Both are printed; the goalposts read the dollar one.

The headline figures

§3 · Was it the path? The stress tests

Is it one era? No. Cut the 7.5-year window in two at 2022-09-30 and the full filter reads PF 0.46 in the first half (39 trades) and 0.68 in the second (39). Both halves losing, so it is not a story about one stretch of one market.

Is it one name? No. In the first episode a single stock was 80% of a whole window’s profit, so this is a house rule now: pull the three largest net contributors (RVMD, ASAN, TEAM, re-ranked from the ledger at build) and the remaining 74 trades read PF 0.29. There was no good filter hiding under a few unlucky trades; the three best names were the only thing holding a bad one up.

Is it the dial? No. The leader cut at 2, 5, 7 and 10 percent; the group cut at the top third, quarter and tenth: 6 of 6 settings of the full filter read under break-even (best 0.97, on 8 and 39 trades; worst 0.50), and the setting that looked best in the first half (full filter, leader 2%) is not the one that looked best in the second (full filter, leader 5%) — the fingerprint of noise, not of a filter.

And the three cells in the wider grid that did touch 1.0 — each paired with its holdout read, never shown alone. leader 5%, no sector rung: PF 1.42 on 106 trades in sample → 0.52 on 120 on the holdout; full filter, sector by the 11 SPDR funds: PF 1.01 on 115 trades in sample → 0.74 on 87 on the holdout; sector only, sector by the 11 SPDR funds: PF 1.01 on 558 trades in sample → 0.86 on 279 on the holdout. Three green cells in, none out. The grid was wide enough to hand us one; the holdout was sealed before the grid was built so that it could not keep it.

Is it the ruler? Partly, on one arm, and we say which. The average trade counts every trade at one dollar risked; the profit factor counts the actual dollars. On the control they disagree (+0.25 R beside PF 0.87) and on the sector-only arm they disagree (+0.38 R beside PF 0.69). We did not chase why they split; both are printed at every rung. On the full filter they agree at every rung, and on the holdout the sector-only arm’s two rulers agree too (-0.39 R, PF 0.41): worse.

§4 · The controls

The same code had to prove it was the same code. In the same in-sample run, after the five arms and before the sealed holdout, the engine also re-took the first episode’s exam: that episode’s own design, its own fixed name list, on the new data store. It reproduced that episode’s recorded figures on 10 of 10 declared metrics across both of that episode’s windows, inside tolerances fixed before the run. So whatever the filter does, it is the filter doing it, not a bad copy of the code.

What the point-in-time universe does, so you know it is not cosmetic. The last column is the same two years with the dead names put back and the survivors-only list replaced by every stock that met the floor on the day: the trade count rises and the average trade falls — the survivorship flattery the first episode itself flagged, now measured. In this lab’s primary window, 38 of the names the control traded later left the market (59 of 783 trades), and every one is in the ledger on the days it was alive. Three guards for names that stop trading were added to the loop; none fired on any arm (0 exits after a name’s last bar, across 5 of 5 arms).

The signature is the control. The median trade sits between -1.02 R and -1.06 R on 33 of 34 cells across both windows (10 of 10 on the five arms; the one exception holds 8 trades). That is the profile the first episode found and the one he describes: most trades lose about what they risked, a few large winners pay for everything. The filter did not change the shape of the distribution; it concentrated the same lottery into fewer, worse tickets.

§5 · The best objection, pre-registered and answered

“You tested a caricature. His hot sector is a judgment, not a SIC code.” Correct, and it is granted above rather than argued with. We tested the rule as he states it on his scans, applied by a machine, because that is the only version a test can reach; a ledger can tell you what the rule does and cannot tell you what the trader does. The two-digit SIC code is coarse — on a typical day only about 13 groups had the 10 members needed to rank at all (about 6 on the holdout), and 6,471 of 19,317 qualifying breakouts sat in a group too small to count, so the sector rung never saw them. A better point-in-time sector map is the first thing that reopens this page. The SPDR-fund variant is the one cut at a coarser map we could run, and it is in the grid above with its holdout read.

“Seven percent is arbitrary.” It is his number, from his own scans, and it was registered before the run. The objection is answered by the dial grid: 2, 5, 7 and 10 percent all read under 1.0 for the full filter, and the one setting that looked good without the sector rung (leader 5%, no sector rung) read 0.52 on the holdout.

“Seven years is one market.” Agreed, which is why the 2010–2015 holdout was declared before the first cell ran and run once: 1,510 sessions on a smaller market (219 members on a typical day). The full filter read PF 0.54 there against the plain entry’s 0.73, and the sector-only arm 0.41. The same direction, on data the filter had never seen.

§6 · Their own terms

Six things on this page run against us. We would rather print them than have them found.

1. The holdout was run once, but its result was not committed-before-run in the way the pre-registration promised. The pre-registration said the holdout would run after every in-sample number was written to the results file and committed. The run log shows one holdout phase and nothing touched after it, and the results write-up was on disk before that phase began (the runner refuses to start the holdout otherwise); what did not happen is the commit of that write-up before the holdout ran. The ONCE half held; the COMMIT-before-holdout half did not, and this page says so rather than implying it.

2. The two rulers disagree on two arms. The control’s average trade is +0.25 R beside a profit factor of 0.87; the sector-only arm’s is +0.38 R beside 0.69. Sizing is not uniform across trades, and we did not chase the split. Nothing on this page reads either arm on one ruler.

3. The sector map is coarse and a third of the breakouts never reached the sector rung. 6,471 of 19,317 qualifying signals sat in groups under the 10-member floor; SIC coverage was 95.1% of member-days. A limit of the mechanisation, stated.

4. Three grid cells touched 1.0 in sample. They are printed with their holdout reads rather than omitted, and none of the three survived. The wider the grid, the more likely a green cell; that is the reason the holdout was sealed first.

5. The regime gate is the ported code, not the one-line summary the pre-registration wrote for it. The summary said SPY and QQQ above their 20- and 50-day averages; the function that ran gates on SPY above its 50- and 200-day averages with a recovery override and cuts slots when QQQ is below its 50-day. The rule to re-use the exact code outranked the summary; the deviation is logged in the pre-registration’s appendix.

6. “Forty-four files” counts the whole results folder. The lab’s results folder holds 44 files; 38 of them are trade ledgers, one row per trade (7,734 rows in all), and the other six are the machine-readable results, the run log and three console captures. Both downloads below are built from the 38.

§7 · The complete ledger — all 2,244 rows