Bagger — what happens if you buy the dip

Win rate on your average cost

Every pick is graded from one cost: your average cost (open + add price) ÷ 2 on the picks whose buy-again zone was entered, and the price you paid — the next day's open — on the picks where it was not. Same picks, same target on the card, one reading. A pick where no add fired keeps the rating it was picked with, so nothing here is scored twice.

Every 100 picks, as dots

The dots are the 1,356 picks shown as 100, coloured by their ending rating. Each rating is rounded to a whole number of dots and the largest rating takes the rounding left over, so the dots always add up to exactly 100; the out of 100 column beside the ratings table is each rating's own independent rounding of its share, so those four figures need not add up to exactly 100.

A dip can only help

91 picks that were not wins become wins (79 were losses, 12 went nowhere). None get worse — a pick that already won stays a win.

No enforced stop is modelled anywhere on this page. The ratings rate every path "ever held". The stop measured from your average cost is reported in Section 4 as a card level only; it never changes a win or loss.

One rating per pick, checked in this order

Each pick gets exactly one of four words. The tests run from the highest down and stop at the first test that passes, so a pick that satisfies two of them is graded by the higher one. The order matters: a pick whose high touched its goal and then fell back well below its cost is a HIT — a win — and not a slip.

1. HITthe high reached the reference area on the card at any point in the window. One day is enough. The test is applied to the day's high and to nothing else, so how the pick closed, and how far it fell afterwards, cannot take this rating away. A win.
2. MOVEDthe high never reached the printed reference area, but one of two movements happened. (a) the ATR test — the pick had a usable ATR and at least a tenth of the days in its window printed a high that cleared the entry by one whole ATR, the same volatility unit the screen's own buy-again zone is measured in; the ATR is the stock's average true range as it stood on the day before the window opened. (b) the 3% test — the pick's highest high in the window was at least three percent above the entry, a flat percentage not scaled to the stock's volatility. Either test is enough on its own. A win.
3. MISSboth of the following, and it takes both. No day cleared the ATR level — and where the ATR was missing or not positive there was no level to clear, which counts as not cleared and which switches that half off entirely — and fewer than a quarter of the days in the window printed a green high — the test is a high at or above the cost the pick is graded on. A loss.
4. NEUTRALeverything left over — the pick failed the loss test above. It never reached the printed reference area and made neither move, but it was not a slip either: a quarter or more of its days printed a green high, or it had an ATR level and at least one day cleared it (just not on the tenth of the days the move day needs). This is no win, no loss — and no win, no loss is not a loss. It sits in the total the win rate is out of and in neither the win count nor the loss count.

The mapping onto win, no win no loss, and loss is exactly this, and nothing else. Win = HIT or MOVED (the win ratings), no win, no loss = NEUTRAL (the no-win-no-loss rating), loss = MISS (the loss rating) — every pick gets exactly one of the four, and the ratings return no fifth word. The win rate this page prints is (HIT + MOVED) ÷ the picks: no win, no loss is in the total it is out of and out of both halves; a loss is in the total it is out of and in the loss count. Two wins are two whichever rating they came from, and a win counts once however far past its goal the price went.

The numbers behind the four ratings

Each test compares against a fixed number. These are those numbers, at the values the ratings actually hold.

The four ratings are defined once, right here. Every table on this page applies these same four words to the one cost that table names — the words are not written out again per table, and no table prints a second cost beside the first.

Every rating uses the same 100 days

We look at the 100 trading days after you bought. A longer window gives the price more time to reach the target, so the same pick can score better over a longer window than over 100 — which is why Section 7 prints each window as its own column instead of mixing them.

Only the day's high counts

Each test reads a day's high — the intraday price — never its close. If the price traded at the goal at any moment inside the window, the pick got there: a touch is a win. That holds even when the pick closed the day far below its goal. MOVED is read the same way: its 3% test is "the price was there", not "closed three percent up". No rating uses the closing price. One later block does count closes — section 7's what-happened-after rows, which include the days a pick closed below its stop — and it is labelled as a count of closes wherever it prints, never as a rating.

The ratings still carry the lows and closes they inherited, so the two can still be compared. The frozen ratings do not read either of them, and a close-based reading is exactly the defect the next card describes.

Four ratings, not five

MARGINAL was dropped. There are four ratings now, not five. An earlier revision scored a fifth outcome from the closing price. It is retired — it appears nowhere on this page, and those four are the only ratings.

Picks with no target are left out, not counted as losses

Where the card withheld the target — the picks with no target on the card, named and re-measured in section 7 — there is no target on the card to grade, so the ratings return nothing for them, and the page counts them out: out of every total it is out of on this page, out of the win rate and out of the loss count together. The same is true of a pick with no usable entry.

Counting them as MISS would have invented a loss: a pick that was never given a goal cannot have failed to reach one. Section 7 puts those picks back on the same target as a labelled what-if, and keeps them out of the ratings while it does it.

Which price every score is measured from

Three different prices could each be called "the entry", and they are not the same price.

Every rating above is measured from one cost per pick. On a pick whose buy-again zone was entered, that cost is your average cost = (open + add price) ÷ 2, where the add price is the lowest price the stock actually traded inside the zone after the add fired, never deeper than the bottom of the zone. On a pick where the zone was never entered, the cost is the price you paid: the next day's open with the modelled slippage applied, rounded to the nearest exchange tick. Neither is the printed entry — the close on the day it was picked, in the card's entry area — and neither is the raw opening price, before costs, which the lanes carry separately and from which the zone is built. Those are three different prices, so the page names which one a figure came from.

The target every rating is tested against is the one target the card carried. The window is the pick's own holding window, counted from the day the add fired on a pick whose zone was entered — the average cost is what those picks are scored from, so the days before it existed are not scored — and from the fill day on a pick where it was not. Each share is therefore taken over the days that pick was actually scored on, and the two never mix inside one number.

Did the price reach the second target too?

Scope: the 932 second-buy picks, over the same dip suffix (add day to the end of the holding window) — a subset of §2, twice over. It holds out the 346 second-buy picks whose card prints no second target, and all 424 picks that never fired an add. That is why its two "reached" rows total less than the ladder's HIT count in §1: 448 here against 1,024 there, the other 246 and 330 HIT picks being the ones those two exclusions remove. This section only records whether the second target was touched too. It is a reach report — it changes no win count.

A positive value means the highest high after the second buy sat that far above the average cost; a negative value means the high never got back above it. The spread is the whole min · median · max, never a single average.

Reach, never a win. Touching the second target adds nothing to the win count — the win count was already set by the target on the card in Section 1. The table only splits the dip-buying picks by how far past the target the path extended.

What we tested

The stop a holder would actually hold after buying the dip is your average cost minus the ATR distance. A rule was tried against that stop: never hold the stop deeper than where the trend broke — replace the stop with the trend-break line max(20-day low, MA20) whenever the line priced higher. On every pick where that replacement bites (the picks the rule changed) the held stop became shallower. Whether a shallower stop gives up paths that were going to reach the goal is a measurable question, and it was measured: each rule-changed pick was walked twice over its own forward days — once on the raw stop, once on the trend-break line — to the same goal, same window, same ordering. The rule is reversed. The measurement, its definitions and its fairness check are below; the card contract it produced is in the box that follows.

On Bagger the choice matters: under pure MA20, 128 second-buy picks have no break below the price you paid at all (freshly-strong names whose MA20 sits above the raw opening price). The trend-break line finds a break strictly more often because it may use the 20-day low when MA20 is above the entry — those are exactly the picks where the invalidation line above the entry is a real chart level rather than an arithmetic one.

Each pick tested twice: the raw stop vs the trend-break stop

One pick, one set of forward days, one goal. The only thing that differs between the two runs is the stop level.

A shallower stop is by construction touched at least as often and no later than a deeper one, so the trend-break stop can never win a pick the raw stop was stopped on. That is arithmetic, not a finding: the counts below are the size of the gap on this screen.

What it cost: wins cut short by the shallower stop

The trend-break stop stopped out the pick, the raw stop did not, and the pick reached the goal — the wins the rule gave up. Printed per stop-rule group; the rule-changed row is the two groups together, never a group on its own.

The ATR spread is the whole min · median · p75 · max of the depth the line removed on those 163 picks, never a single average. Every pick in the cost cell is a win (HIT or MOVED) on the dip-buying ratings — the same picks those ratings already count as wins.

What if your stop were the trend-break line instead of the raw ATR stop?

How often each stop was touched before the goal, on all 932 second-buy picks and split by stop-rule group. Each run applies its own stop to the same picks — the two rates share one total (the picks in this group column), so the last column — the count of picks the trend-break stop hit that the raw stop did not — is the honest comparison.

Every rate above is a stop count, and these picks contain winners. The four ratings enforce no stop at all, so a pick stopped out and then reaching the goal is still a win (a HIT or a MOVED) — the two counts overlap and neither is readable without the other. Split by ending rating, the same three sets read: the rule-changed set — 606 — HIT 456 / MOVED 127 / NEUTRAL 10 / MISS 13 — 583 of 606 are wins (96.2%); the picks the raw stop hit — 252 — HIT 104 / MOVED 125 / NEUTRAL 10 / MISS 13 — 229 of 252 are wins (90.87%); the picks the trend-break stop hit — 417 — HIT 267 / MOVED 127 / NEUTRAL 10 / MISS 13 — 394 of 417 are wins (94.48%). Two rows are counted over a set those three do not cover: the break_below_blend group — 179 — HIT 136 / MOVED 38 / NEUTRAL 1 / MISS 4 — 174 of 179 are wins (97.21%); all second-buy picks — 932 — HIT 694 / MOVED 200 / NEUTRAL 13 / MISS 25 — 894 of 932 are wins (95.92%).

Outside the rule-changed set the two stops are the same on two of the groups (no_break, break_below_blend — the trend-break line is drawn at or below the raw stop, so the rule changes nothing) and differ on the third, break_above_band — the whole zone is drawn below the trend-break line, so the rule-changed stop is the line and the raw stop is not. That group holds 144 picks here and is the whole of the gap between the rule-changed delta and the all-second-buy delta. The table is additive: the rule-changed subtotal row is its two groups, and the last row is all four groups.

After the stop was hit: did the price come back?

Restricted to the 417 rule-changed picks the trend-break stop hit. Read on a day strictly after the stop-out day, through the window end.

Counting the stop-out day itself as well: 397 of 417 traded back to ≥ the average cost and 267 of 417 reached the goal. Of the 267 that reached the goal, 163 are picks the raw stop won (the cost cell above), 104 are picks the raw stop's deeper stop was touched on too — so the win was given up on both stops — and 0 are picks where neither stop won.

A stop that is hit is not evidence the pick was a loss: the four ratings for the same pick are unchanged, and they enforce no stop at all. This table counts where the price went after the stop was touched, nothing more. Split by ending rating, the two sets it rates read: the rule-changed set — 606 — HIT 456 / MOVED 127 / NEUTRAL 10 / MISS 13 — 583 of 606 are wins (96.2%); the picks the trend-break stop hit — 417 — HIT 267 / MOVED 127 / NEUTRAL 10 / MISS 13 — 394 of 417 are wins (94.48%).

Was the test fair? Did the second buy happen where the rule allows?

The dip lane credits the second buy at the realized dip low inside the buy-again zone. On a break_inside_band pick the trend-break line is drawn inside the zone, so that low can sit below the line: the buy then prints at a price the guard forbids and the average cost lands below the very stop meant to cap it. On a clean_cap pick the line is drawn at or below the bottom of the zone, so the buy is always at or below the line and the problem cannot arise.

The check. The measurement is repeated with the buy repaired into the valid part of the zone — buy price = max(dip low, trend-break line) — on both runs, so both sides are priced off the same buy. On the 256 clean_cap picks the repair is the identity (0 buys moved) and each run comes out identical. Across the whole rule-changed set the trend-break run is unchanged on 0 picks: the repair moves the raw run only, and it moves that run's outcome on 10 of the 606 rule-changed picks (it moves the credited buy price on 258 of them). The trend-break stop is max(repaired raw stop, trend-break line) and raising the buy never lifts the raw stop above the line on these picks, so it still resolves to the line.

What the repair costs the cost cell. On break_inside_band the cell goes 128 → 118 — the repair removes 10 of the 128; on clean_cap it removes 0 of 35. Read on the repaired buy on both sides the rule-changed cell is 153 of the 606 rule-changed picks, against 163 as measured. On the picks the repair removes, the raw run had been handed a buy below the line — its raw stop sat shallower than a valid buy allows and it registered a win the rule-consistent raw stop does not; the trend-break run stops out either way. It is the trend-break line, not the buy price, that stopped the position.

Rounding left over on the as-measured picks: on 43 trend-break-stopped picks the trend-break stop sat strictly above the average cost — the position was bought at or below the level that then stopped it. After the repair that count is 0. Of the 417 trend-break stops, 117 land on the add day t0 itself — the day where the credited buy and a touch of the trend-break line coincide.

Does where the trend-break line sits change how a pick ends?

Chosen line: the trend-break line, max(20-day low, MA20). The group does not separate wins measured from your average cost from the rest on this yardstick — the win share is broadly flat across the groups, and the non-wins end with the same rating in every group because the four ratings enforce no stop.

Pure-MA20 variant, second-buy picks (for contrast)

What the group tables do and do not say

The four ratings never enforce a stop. They grade every path "ever held" — the room the price gave — so the stop-rule groups never change a win or loss. What a group says is only where the trend-break line is drawn relative to the stop measured from your average cost and the buy-again zone.

That is exactly what the two-stop measurement above supplies and the ratings cannot: the ratings cannot count the paths a stop would have ended, because they model no stop. Walking each rule-changed pick twice on its own days is what turns the rule into a countable cost, and it is why the rule now ships as the invalidation line rather than as the stop.

Picks with no second buy (424) never added, so they have no stop measured from the average cost to cap; their stop-rule group is informational geometry of the would-be buy-again zone and is not part of the rule-changed numbers above.

One pick day = one screen run

The ratings, grouped by the day each pick appeared, on the one cost this page grades from — 1,302 of 1,356 picks win (96.02%). Each day's win rate = (HIT ∪ MOVED) ÷ that day's picks; no win, no loss is in the total it is out of, not the win count. Days are small (Bagger screens range from 1 to 12 picks, median 3), so a day rate is a descriptive number on a tiny sample — not a ranking rule, and it says nothing about the next day.

Extreme days, and what they shared (descriptive — tiny n, do not rank)

Config columns. grade = the model's own S/A/B mark for the day's picks; "share graded S" = the percent of the day's picks marked S; Final_Score is the pick's 0–1 score (higher is better) and Rank is its position in that day's ranked list (1 = the top) — the two inputs the mark is built from; flow_state is the flow-chip state (accumulate / distribute / retail) on that screen; strategy_class is the chart-setup class. Every day on both screens joined at 100% to the mark source, and flow / strategy coverage is also 100% across the picks.

The best days and the worst days share these config values: grade B; flow-chip mode accum_retail_sell; strategy class breakout_momentum / none — which is the point of the "do not rank" warning: the extremes do not separate on config.

One cost, one stop

Every pick on this page is graded from one cost: where the second buy fired, the blended cost — the average of what you paid for the two fills; where the band was never entered, the next day's open, because that is the only fill there was. Each pick has one stop on that same path: the average-cost stop where the second buy fired, the entry stop where it did not. They are never the same price and never share one word.

234 of 1,356 picks (17.3%) reached that stop before they won. That is the headline of this section. It is not the stop-out rate — the stop was reached, which is not the same as the pick ending in a loss, and the table below says what each one became.

How the 234 are arrived at

Each step is a partition, not a subtraction chain read left to right: every pick is in exactly one of the three buckets.

1,356 picks → 819 reached the stop at some point (537 never came near it) → 585 of those had already won, and a win is a win → 234 counted in this section.

Why the 585 leave. Those picks reached HIT or MOVED — a win — before the stop was ever reached. The stop never got there first, so counting them would be counting a win as a stop problem. This is where a stop count usually goes wrong: a pick that won and then drifted back to the stop is still a win, and every one of them is here, on its own line, rather than folded into the total.

How far it went, and how far it came back

For the 234: at any point in its window, the median pick's lowest low sat 10.23% below the stop it was supposed to cut at, and the median pick afterwards reached +19.97% above the cost it was bought at. 4 of 234 never traded back above that cost at all. Both figures are medians of each pick's own extreme, measured against the level named in the column — the stop for the depth, the cost for the recovery — so a row is reproducible from its own definition.

How far below the stop each rung went

The same excursion, split by the rung the pick ended on — so the rungs that lost are in the table beside the rungs that came back, rather than only the winners being measured. Every column here is taken from the stop's touch through the pick's own ending: the bar the pick's win was established on where there was one, the end of the window where there was not. That is a shorter window than the whole-window figure in the paragraph above, so the two do not have to agree — they are one event measured over two spans, and each row here answers only for its own.

Depth is a share of that pick's own stop, so a row is reproducible from its own definition (one cost, one stop: the average-cost stop where the second buy fired, the entry stop where it did not). Least deep is the pick in that rung that barely dipped, deepest the one that went furthest; typical is the median and average is the mean, both printed because a single pick that dived moves the mean and not the median — on this deck's MISS rung the median is 32.09% while the deepest of the same 33 is 78.47%. The shallowest rung is HIT at a median of 5.46%. The last column is the same picks' highest high after the touch, against the cost they paid — the rebound that decides which rung the row is in.

What the rebound is and is not. A rebound to HIT or MOVED is a real win by this page's definition and is counted as one — but it is a win that needed the stop to be held through. Whether holding is worth it is a question about the holder, not about the pick, and this section does not answer it. It reports what happened to the picks the stop was reached on, before they won. It is not the stop-out rate. The stop-out rate is the other two rows of the table: the 46 of 1,356 (3.4%) that reached the stop first and ended a scratch or a loss. The 188 that came back to win are the reason the two numbers differ, and reading this count as a loss rate is the error the exclusion above exists to stop. It is not the §2 ladder either: the ladder grades the pick, this grades the path to the stop.

Where these numbers come from. A lane that has to reproduce a reviewed number before it may report a new one: it re-derives the buying-once deck and lands on that deck's known cut-before-win tally of 310 of 1,211. Failing that it reports nothing. The win's own time is the meter's establishing event, never a close: HIT at the first bar whose high reaches the target, MOVED at the bar the meter's own test first holds.

What it shows: the same four ratings, re-counted on the blended cost this page grades from, inside each mood: win rate (HIT ∪ MOVED) was 96.7% on bull-condition days (n 1,012), 94.3% on mixed days (n 157), and 93.6% on bear days (n 187) — a spread of 3.1 points from bull to bear, on the rounded rates printed above. That is a same-day pattern, not a forecast, and the smallest group wobbles most.

Which basis this is, and which it is not. The companion pages carry a mood card too, and it is not this one: it re-counts at each pick's best entry, this re-counts on the blended cost, so the two pages print different rates for the same three groups and neither may be read as the other. The group sizes are the deck's own and so are identical on both — 1,012 / 157 / 187, summing to the 1,356 picks this page grades. Pooled back together the three groups reproduce the section 1 headline exactly, which is the check that keeps this table from contradicting it.

What these counts cover

The win-rate sections above answer did the pick win. This one does not, and neither does the section directly above it — section 6 asks which picks reached the stop before they won, and this section asks what did the price do on the way: where the worst drop landed, how long it sat below the stop, whether the buy-again zone entry came before the stop.

Which price it is measured from. This section is on the actually-paid, buying-once measure only — the entry is the next day's open, which is the one cost every figure below is graded from — and the goal is the target on the card, over the pick's own holding window (100 days, day 0 = the day you bought). The headline at the top of this page is a two-cost composite, and nothing in this section is graded on your average cost: nothing here is a result of that headline, and nothing here is printed beside a rate.

Which picks are counted. 1,356 picks were measured — the picks whose card carried a target. A further 714 picks have no target on the card: their card withheld the target and no price path was recorded for them, so they sit outside every count in the first two tables below. They are not unmeasurable — the third table puts them back on the same target.

Day numbering, and why no figure here carries an offset. Every cell below is a count of picks, not a day, so the day-numbering rule moves nothing on this page — it is stated because the two families are read differently underneath and reading one as the other is the defect. The three fields whose name begins day_of_ hold an absolute position in the window, and a position is written 1-based so that day 1 is the fill day. The six duration fields — the fields whose name carries days — hold a duration and take no such offset, so a duration of 0 is a real value meaning "on the day it was counted from" — days_below_stop of 0 means the pick never closed below the stop after the first touch, never a missing day.

Counts, not rates. Every cell states the set it is measured on, because a count whose total is invisible cannot be reconciled by a reader; a count equal to its own set means the field is defined on every pick of it. The sequence fields are counts too: late_zone_entry is a rare tail and is printed as counts, never as rates.

The counts above whose numerator is a stop-touch event print their own ending partition beside the count, because most of the picks inside them had already reached the goal first. A pick that got to its goal is a win however far it fell afterwards — the ratings defined at the top of this page are scored on touches, and a touch is a win — so a count of stop touches is not a count of losses. On this screen it is the other way round: 541 of the 561 picks that closed back at or above the entry after the buy-again zone and then the stop had already won. The remaining stop-touch picks name in their own labels what they count — a day position, a sequence, a later goal touch — or are the second spelling of a pick that already carries its partition.

The buy-again zone sequence is only applicable where the zone top and the stop are both on the card. It is applicable on 1,322 of the 1,356 measured picks; on the rest the sequence reads false because there is no zone to enter, which is an answer, not a missing one. The stop-side sequence (stop_then_recovered_to_green) is counted from the stop day onward, so it is silent about whether the pick was green first — section 6 carries that distinction and measures the narrower set.

The same ratings over a longer window (counts only)

The three longer windows re-run the same ratings on the same measured picks, at the same target and measured from the same price. They are a longer look, not a better rule — and they override the time-stop.

Windows. The declared-window row is the same ratings themselves on the measured picks: its winners are 1,211 of 1,356 — the same figure this section's top row prints, on the executed open and not on the masthead's blended cost. The +10, +20 and +50 rows are the same basis at a longer window, not a fourth basis — the target and the entry are unchanged, only the days watched — and because a longer window override the time-stop, every longer-window column prints as rating counts only: no share is printed for it, and none is authorized. Rescued = picks whose declared-window rating is a loss and whose longer-window rating is a win, again a count only.

Same ratings, same target, two sets of picks

The picks whose card withheld the target have stood outside every printed rate. They are not unmeasurable — only differently measurable: the source data carries the target the card withheld, so the same ratings can be re-run over it.

Which total. The overall figure and the two set figures are not interchangeable. The overall row's total (2,070) is new: it re-weights the same ratings toward the no-target picks, which are 34.49% of the file, so both totals are shown and a reader can see which one a figure came from. The top row is this section's own buying-once figure — every pick graded from the executed next-day open, which is this section's basis and not the blended cost the rest of this page grades from — 1,211 of 1,356 = 0.8931. It is not the figure the masthead prints: the masthead grades from the blended cost, the average of your two fills, so the two counts differ because they price the same picks from different money.

The one label this row may not go without. The no-target share — 95.38% — answers if the ATR target had been applied. On a no-target pick the card withholds the target (it nulls the printed target and keeps the stop), so no target was on the card: this is a what-if share, not a real rate, and it is printed only beside the buying-once and overall rows, each with its own total. Nothing here re-grades the ratings — on the buying-once measure these picks are still out, which is what the scope note at the top of this page already says of them. The same 714 picks are the ones the section 6 ledger names and stands aside, and they are measured the same way any pick this screen printed is, but the reason they need re-measuring is worth stating exactly: the card withheld one number — the target. It printed the entry and it kept the stop, which is why those 714 picks carry a hard stop like any other pick. What is missing is not a card field but a recorded row: b2's no-target branch records no price path at all, so the entry, the target and the stop are absent from that lane together — which is also why the what-happened-after counts in the table above stop at the 1,356 measured picks and are not extended to these 714. The source data supplies all three — the entry and the target from the control table, the stop from the card's own hard stop — and then re-runs the identical ratings over the same target.

One family of rows has no answer on the no-target set. The buy-again zone needs a target on the card, so the zone top and the zone bottom were absent on every no-target pick and the zone sequence is not applicable there — a fact about the numbers behind it, not a zero. The zone rows (day_of_zone_entry, days_in_zone, zone_recovery_days, zone_stop_then_recovered, zone_stop_recovery_days, late_zone_entry) therefore have no answer on that set, and a 0 printed there would be a share over an empty total. The stop family does apply to it: a no-target pick carries the same hard stop as any other pick — its card withheld the target only — which is why the lane measures the stop family on it and records no value at all for the zone rows.

What this section is, and which fill it is graded from

The sections above grade one fill price. This one grades a different one, and it says so with both numbers before it prints a single band.

Which fill. Every rate below is on the executed next-day open — the price you actually paid, the same basis section 7 is on — and not on the average cost the masthead grades from. The two are not close and neither may be read as the other: this section's 620 of 733 = 84.6%, against the masthead's 1,302 of 1,356 = 96.0%. The gap is the add-on-the-dip cost doing its work, not a disagreement about the picks.

Which picks. The tape starts on 2025-09-01 and the flow measure reads a 21-session window, so a pick scanned before 2025-10-01 has no full window behind it. 623 of the 1,356 picks are out for that reason — a calendar warm-up, not a missing-data rate — and the bands below are read on the remaining 733. The cut costs most of the deck because the screen has run for years and the tape has not: it is the tape that is short, not the screen.

The six bands

Every table below groups the same 733 picks, so the All picks row is the same row in all six — and every share is HIT ∪ MOVED over the picks in its own row.

How to read a band against the row under it. The bands are the whole story, unaggregated: read each row against its own All picks row and nothing is hidden in a summary. A row of fewer than 30 picks is marked ⚠ on its own count and is not read — with this many bands some will look good by chance, which is why the verdict below never rests on one row.

What the flow measure covers, and what it does not

The convention is named because two are available here and only one of them is honest.

The convention: a value, not a non-empty cell. Every band column below is 100% populated — and that is exactly why "populated" cannot be the measure. A pick whose band column reads no data has no value there; counting it as covered would print 100.00% coverage for sm_cleared_band, which publishes nothing on 195 of these 733 picks. no data is a row in the tables above, not a hole in them, and the count below is of picks that actually carry a value.

What a blank means — and what it does not. The cleared columns ask what share of the peak's buying has since been sold back out. That ratio needs both of its phases to carry at least 5 observed flow sessions before it means anything, and the measure publishes no number at all when they do not — never a zero. A zero would read as "nothing has cleared", which is a different and stronger claim than "there is not enough tape here to say", and the band's no data is the honest of the two. The same 5-session floor is what makes a blank in the source column line up one for one with no data in the band beside it, on all 733 picks.

Columns this section does not print, and their gaps. Stating these is the point: their absence here is not a claim that they are complete. Each is a gap, stated as one: the peak fields (peak_date, peak_drop, peak_days) and exit_days — no value on 111 of the 733 picks. · first_close_over — no value on 73 of the 733 picks. · time_to_tp — no value on 241 of the 733 picks. · rt__extremes — no value on 733 of the 733 picks — it carries the literal No data on every one, so its coverage is 0%, not 100%.

What the bands say

The direction is stated before the read, so a sign that comes out the other way is a finding rather than a reframing of the question.

The hypothesis. Bandarmology puts accumulation before markup: a pick whose visible buyers were accumulating should win more than one whose visible buyers were distributing. Every component below is read on that direction.

0 of 6 components separate in the hypothesised direction. 3 of the 6 hold the same sign in both screens, and 1 of those — sm_consistency (-5.4pp) — runs the other way in both. A component that does not separate, or separates backwards, is a result and is printed in full above rather than dropped.

Why a big gap is not the finding. sm_cleared reads +21.1pp on bagger and -3.2pp on swing; sm_trend reads +17.5pp on bagger and -3.9pp on swing; sm_avg_trend reads +15.2pp on bagger and -3.9pp on swing. A component whose sign flips between the two screens is not read as a pattern, however large either gap is. The rule that rules one out is the rule that would have ruled it in had the sign held: a cell is a pattern when it is ordered, not when it is a single hot slice. The two rules below this one — a spread under 10.0pp is not read, and a slice under 30 picks is not read — are marked on the tables above.

The model these picks are the record of

These frames are one model's historic picks. Read "the model that chose these picks", never "the model in use now" — a retrain does not rewrite history, and this page is history.

Every pick on this page was chosen by the bagger model, registered 2026-09-04. The chain runs the screen that chose the picks (2026-09-05) → the chosen-set deck (2026-09-06) → the frame this section reads (2026-09-12). The retrain — a later model on a wider name list — was registered 2026-09-13, after every one of them. It produced none of this history, so nothing above is a measurement of it.

How that is known rather than assumed. The screen that chose these picks logged, on every row it gated, the threshold it applied. Grouped by rating level that threshold is one number per level, and those are exactly the numbers registered against the model above — the values and the row count are withheld from this public copy. A different model would put different numbers there, so the vintage is shown by a figure that would move if it were wrong, not asserted in prose.