The technology behind the CRWD trade
Adaptive Mean + σ Bands v5.2 — one indicator, three zooms, one scoreboard.
This is the tool that called the quiet on CRWD on 7/28, gave the trigger on 7/31, and confirmed on 8/3.
This page is the honest manual for it: what it does, what its scoreboard really says, what it
cannot tell you yet, and what to fix next.
Build
v5.2
shipped 8/3/2026
Zooms
3
daily · 30-min · 2-min
Trade types tracked
9
separate scoreboards
Status
HYPOTHESIS
not a validated $-edge yet
Read this before you quote a number off it
The scoreboard inside the indicator is the
tool's own tally on the chart you have open.
It is not a backtest. It has no commissions, no slippage, and it was tuned on the same charts it is
scoring. Treat every number on it as
a question to go test, not an answer.
Every chart has a "normal price" and a sense of how far price usually wanders from it. The tool picks
the right normal for the chart you're on, then draws the wander as bands.
| Chart | The middle line it uses | What that chart is for |
| Daily & up | 20-day average | WHERE — is this a place to be interested at all? |
| 30-min | true 5-day average | WHEN — the trigger and the stop. The swing lives here. |
| 1–29 min | session VWAP + the 50-minute line | HOW — timing and tape only. Day-trade stops. |
1) It spots the quiet
When the bands squeeze into the tightest they've been in about a month, something is loading.
Quiet at the middle line = a coil (can break either way). Quiet far from
the middle line = a base (a pause after a move that already happened).
2) It marks the exact bar things change
Quiet → breaking out → stretched. Each flip gets a small tag on the candle where it happened, so
a state is an event you can point at, not a feeling.
3) "Stretched" has to be earned
This is the most important rule in the tool. When a range breaks, that range projects a
target — the range height added to the edge. Until price reaches that projected target, the tool
refuses to say "overextended." A push inside the projected move is trend, not a top.
Nothing yells "sell, it's stretched" while the move is still owed.
4) It takes its own trades, with real stops
Stops come from structure — the hold low, the base's far side — not from a feel number. One position
at a time. Entry, stop and target are drawn on the chart from the entry bar and frozen at the exit
(that's what v5.2 added), so a screenshot shows plan vs. outcome.
5) It keeps a scoreboard per setup
Nine setups, scored separately: VWAP pull-in long, expansion long/short, continuation long/short,
base-break long/short, backside long/short. The aggregate can't tell you which setup works. This can —
with the caveats on the next page.
The slide said: VWAP L: +.48 ExpR over 168 signals, 26% win
That line does not survive a question from the room, for two separate reasons. Both are fixable.
Problem 1) "26% win" is not the win rate
Look at how the tool fills that column. A trade only counts as a W if it exits by
hitting the full target. It only counts as an L if it exits by
hitting the stop. But every trade counts in the total — including the ones that ended by
trailing out, getting covered at the opposite band, being flipped by a reverse signal, or timing out.
Those land in the total and in neither column.
So 26% is the target-hit rate. The other 74% are not losses. And you can prove that
without any new data, straight off the two numbers already on the slide:
target ≈ 2R, hit by 26% of trades → 0.26 × 2.00R = +0.52R
reported average per trade = +0.48R
therefore the other 74% contribute 0.48 − 0.52 = −0.04R in total
−0.04 ÷ 0.74 = −0.05R each, on average
if those 74% had really been −1R losses:
0.52 − 0.74 = −0.22R per trade, not +0.48R
The honest read
26% of them ran to the full target. The other 74% averaged about
−0.05R — near flat,
not −1R. It is not a 26%-win strategy. Read the
W-L cell on the table for the real
split; never quote the Win% cell on its own.
Problem 2) Those 168 are not "this exact setup"
The deck says this exact setup has paid +.48R across 168 of them. The 168 is
every VWAP pull-in long the tool took on that chart — every day, every regime, no filter.
The CRWD trade had four extra things going for it that those 168 do not share:
- a coil — the tightest bands in a month, flagged 7/28
- a PDH break — the day's first break of yesterday's high
- a 50-DMA pullback behind it — the reason the whole thing was on the radar
- all three timeframes agreeing
So 168 is the sample for the entry mechanic, not for the setup. The
filtered sample — coil + PDH + aligned — is much smaller and has never been counted. That's the real number,
and we don't have it yet.
Three more things that shrink what 168 is worth: one symbol on one chart (in-sample); the tool only holds
one position at a time, so signals that arrived while it was busy were skipped — the 168
are "signals that got a slot," not "signals that fired"; and the tool's settings were tuned across
v1.0 → v5.2 on these same charts.
Fix 1 — "Intermediate Time Frame" · Bucket edge card
VWAP L: +.48 ExpR over 168 signals, 26% win
VWAP L, engine tally on this chart: +0.48R avg over 168 fills. The 26% is the
target-hit rate, not the win rate — the rest exit by trail/cover, averaging about −0.05R. Unfiltered:
no coil, PDH or alignment gate. Hypothesis, not a proven edge.
Fix 2 — "Daily Time Frame" · Bucket edge card
EXP L: +.24R over 9
EXP L: +0.24R avg over 9 fills — n=9, far too few to judge. The tool greys out
anything under 10 for exactly this reason. Context only.
Fix 3 — "Technology" and "Trade Review" slides
This exact setup has paid +.48R per trade across 168 of them.
The entry mechanic — the VWAP pull-in long — has averaged +0.48R across 168 fills on
this chart. That is the mechanic, not this setup: those 168 had no coil, no PDH break and no
timeframe-alignment filter, and CRWD had all three. Getting the filtered number is the next job.
Traded the timeframe with the proven edge: this setup pays +.48R a signal over 168 of them
Traded the timeframe that gave the trigger, and took the stop from that same chart.
The +0.48R tally is the reason the mechanic is worth testing — it is not proof, and it did not size this trade.
If someone in the room asks "26% win — why would you trade that?"
"Because 26% isn't the win rate — it's how often it ran all the way to target. The tool only scores a W
at the full target and an L at the stop; everything trailed or covered lands in neither. If those 74%
really were losses the average would be minus 0.22R, not plus 0.48R. And I'm not claiming it's proven —
that's the tool's own tally on one chart. The filtered version is what I'm building next."
The tool exits at its target. So it has nothing at all to say about
whether you should have held past the target.
The single biggest line in the CRWD review is "banked 2R while the trend paid 6R." That is a question
about the runner. But every one of those 168 trades was closed at target, stop, trail,
cover or timeout — none of them was ever allowed to run past the objective. The scoreboard measures
getting in. It has never measured staying in.
Don't put the +0.48R number next to "I should have held for 6R"
They're answering different questions. The stat supports taking the entry. It says nothing about the hold.
Putting them side by side on one slide is what makes the room think the stat backs the hold rule. It doesn't.
The three hold rules that need testing
| Rule | What it does on CRWD | Tested? |
| Take the full 2R and be done | +2.00R at 197.68 — what actually happened | this is the baseline |
| Trail behind the 50-minute line past 2R | would still be in as of 8/4 | no |
| Hold to the projected move, then manage | objective 209.15, tagged 8/4 | no |
Until those three are run over every historical firing, "hold longer" is a feeling with a good anecdote
behind it. One trade that would have paid 6R does not make a rule — the same rule has to be checked against
every time it would have given back a 2R winner.
Same skeleton, three months apart. Big MA holds the pullback → the 5-DMA reclaim is the
trigger → momentum burst.
| NVDA · May 2026 | CRWD · Jul–Aug 2026 |
| Pullback into | rising 20 DMA ≈ 199–200 | 50 DMA — 7/29 closed 179.38 vs 50 DMA 179.80 |
| Trigger | 5-DMA reclaim 5/6–5/7 | 5-DMA reclaim 7/30 (185.22 vs 181.96) |
| Then | ran to 225.83 — ~+38% off the base | ran to 212.64 by 8/4 — +18.7% in 4 sessions |
| The failure | managed a daily swing on the 1–2 min. 12+ round-trips on 5/4, −$59.98 | banked 2R at 197.68; trend paid 6.6R. Skipped the 8/3 re-add signal |
What NVDA taught, that showed up in CRWD
- Don't buy the touch — buy the reclaim. The early CRWD dip-buy at 191.15 stopped −1R.
That was the knife. The real entry came after the base formed: coil → PDH break → VWAP hold.
- The stop comes from the chart that gave you the entry. 30-min gave the trigger, so the
stop was the 30-min hold low at 187.98 — structure, not ATR, and not the 2-min.
- Intraday cooldown is not a state change. This is the NVDA lesson literally turned into
code: the "stretched must be earned" rule was written into the tool so it can't shout overextended
while the projected move is still owed. NVDA cost $59.98 to learn; CRWD got it for free from the rule.
- Make the timeframes agree before sizing. Daily quiet 7/28 → 30-min trigger 7/31 →
daily confirm 8/3.
What did NOT carry: the hold
NVDA's mistake was managing a multi-day hold too tight and giving it back in round-trips.
CRWD is the
same disease at a smaller dose: 2R banked, 4.6R left, re-add signal skipped.
And the fills say the same thing again — 7/31 overnight +$68.36 against day-session −$6.16; 8/3 −$11.09
across 39 fills at 44%.
The hold made the money. The churn gave it back.
Three months, two names, one unfixed habit.
8/4 close
211.22
high 212.64 · +6.63R off 191.21
Projected move
209.15 ✓
tagged 8/4 — paid
Daily +2σ
211.87
close sits at +1.93σ
Daily +3σ
221.31
the NVDA-style burst target
The rule that let you hold just expired
You held through the run because the projected move was still owed — that was the tool's rule and it was
right. On 8/4 price tagged
209.15.
The move has now been paid.
From this bar on, the daily "overextended" flag is armed: one close above
211.87 and the
tool flips to OVEREXTENDED UP —
no new longs, no adds, don't chase.
This is no longer a hold-the-thesis trade. It is a trail-management trade. Those are different jobs.
The levels that matter now
| Level | Price | What it means |
| Daily +3σ | 221.31 | Where a real momo burst would stall. Hypothesis only — CRWD hasn't done it. |
| Daily +2σ | 211.87 | Close above = stop adding. Trim zone, not entry zone. |
| Projected move | 209.15 | Paid. The reason-to-hold is used up. |
| 5 DMA | 193.84 | Rising fast. The runner's line. |
| 20 DMA | 193.24 | Reclaimed 8/3. Loss of this = the swing is over. |
| 50 DMA | 183.20 | Where the whole setup started. Nowhere near it. |
Manage-from-here — written down BEFORE the next bar
- The core: trail under the rising 5 DMA on a daily close basis. Not intraday.
That is the NVDA doctrine — the 1–2 min never drives the exit on a swing.
- No new adds above 211.87. The projected move is paid; the tool's own gate has armed.
Adding here is chasing, and chasing is what the 8/3 fills already punished (−$11.09 on 39 fills).
- 221.31 (+3σ) is the trim objective for the runner, not a prediction. If it gets there,
that's where the NVDA analog says momentum burns out.
- The swing is done on a daily close back below the 20 DMA (193.24). That's the line,
and it's far — which is the point of a trailing rule instead of a tight one.
- No CRWD day-trading around the core. Both sessions of scalping lost money while the
hold made it. If the intraday edge is gone — under 45% after 10 fills — stand the name down for the day.
Right now this trade was captured by hand, after the fact. That's why the
deck took a night. The pipes to do it automatically already exist — the indicator just isn't plugged into them.
The gap
The indicator ends with 16 alertcondition() lines. Those can only send a
fixed sentence — "Play 1 — held session VWAP, box break up". They cannot carry the price,
the stop, the target, the size or the bar time. So nothing about this trade could ever reach the database.
The live pipe wants the other Pine function — alert() — which can send a JSON payload.
It's already running for the other indicators:
Pine alert("AMSB|{...}")
→ webhooks/tradingview_smb_indicator_trigger.php?key=…
→ Postgres alerts_tradingview.indicator_alerts
→ the scorecards, the trade-check page, the playbook builder
The fix — one block added to the indicator
On the entry bar, fire an ENTRY with the plan; on the exit bar, fire an EXIT with the reason. The receiver
requires four fields (sym, tf, event_type, indicator) and
keeps every extra key in raw_payload, so nothing is lost.
// on the entry bar
alert('AMSB|{"indicator":"AMSB","sym":"' + syminfo.ticker + '","tf":"' + timeframe.period +
'","event_type":"ENTRY","direction":"' + (tDir==1?"LONG":"SHORT") +
'","entry_px":' + str.tostring(tEntry) + ',"stop_px":' + str.tostring(tStop) +
',"tgt_px":' + str.tostring(tTgt) + ',"qty":' + str.tostring(lblQty) +
',"setup":"' + tKind + '","ts":' + str.tostring(time,"#") + '}',
alert.freq_once_per_bar_close)
Then the same shape on exit with "event_type":"EXIT", exit_px,
reason (STOP / TARGET / WAVE5 / TIME) and the R multiple.
Carry a trade_id on both legs so they pair up.
What that buys you
- Every firing lands in the database the second it happens — that's the filtered sample the +0.48R
number is missing.
- Your real fills can be lined up against the signals automatically — the trade-check the deck
does by hand.
- The annotated chart builds itself:
SetupChart.open() already takes bars plus
entry/stop/target lines and markers. Feed it from the same rows.
- Next deck starts with data instead of a blank slide.
Order of operations — your own rule
Don't patch the Pine on sight. The rule is:
test the logic in the a2 engine first, then migrate
to Pine and diff. The alert block is plumbing and is safe. Any change to entry/exit
logic goes through the backtest first.
v5.2 is a day-trading tool that can also swing. v6.0-swing is only the swing.
Everything that wasn't the NVDA or CRWD trade has been cut, and the whole thing now records itself.
→ adaptive_mean_swing_v6.pine
— paste into a new TradingView indicator.
Cut
| Gone | Why |
| The whole backside / fade layer — plus base-break trades, the deep-extension and double-tap filters | Fades are day-trades. Neither playbook is one. |
| The 1–29 min execution frame and the 50-minute rail + state cloud | A swing isn't executed on the 2-min. Session VWAP stays — Play 1 needs it on the 30-min. |
| Continuation setups (weakness-after-strength) | A different trade. Not the one that paid. |
| 2-day / WTD / MTD vwaps | Chart furniture on a swing frame. |
Added
| New | What it fixes |
| RECLAIM entry — pullback into the big MA, then close back above the fast MA | The NVDA trade had no code at all. It's how the $164→$225 leg started and how CRWD turned on 7/30. The tool could not take it. |
| Projected-move target | Both playbooks aimed at the objective, not an R multiple. Now the default. |
| "Trail the fast MA AFTER the move is paid" | The CRWD forward rule, in code: hold while the move is owed, trail once it's tagged. |
| alert() JSON on entry + exit, paired by trade_id | The capture pipe. v5.2 could not send a single price to the database. |
Two real bugs found in v5.2 while stripping it
- Max bars in trade = 26. On a 30-min chart that force-flattens a swing after about
two days. The CRWD trade survived it with a few bars to spare; an NVDA-length hold would have been
thrown out mid-trend by a timer. Now 500.
- The edge table's Win% is what put "26% win" on your slide. A trade scored a W only
at the target and an L only at the stop — trailed, covered and timed-out exits counted in the total and
in neither column. Now W-L-S, plus a separate Prft% for any exit above zero.
The bad number came from the tool, so the fix belongs in the tool.
Not yet tested
This build compiles as written but has
not been run against history, and the RECLAIM
entry has never fired anywhere. Per your own rule — logic gets tested in the a2 engine before it gets
trusted. Load it on CRWD and NVDA first and check it reproduces the two known trades before it goes
anywhere near an automated alert.
Every rule in this tool came from getting something wrong first. That's the part worth showing the room.
| v | What went wrong | What got built |
| 2.2 | Continuation trades almost never fired — a fade was always holding the one slot | Trend signals now push a fade out of the way and take the slot |
| 2.3 | Rode a trend short all the way back to the mean | Cover at the opposite 2σ re-entry — the impulse is done |
| 3.9 | Churned long/short around a flat VWAP | A pull-in needs a leg to pull in from |
| 4.2 | On a trend day price never comes back to VWAP, so re-entries dried up | The 50-minute line as the pull-in rail |
| 4.5 | The NVDA mistake — "stretched" fired mid-move and scared out the hold | Stretched must be earned — no overextended call until the projected move is paid |
| 4.9 | Faded the first poke out of a fresh base | Fades inherit the same earned-extension gate |
| 5.0 | Overnight leaked into the prior day's VWAP | Anchor at the first bar of the trading day |
| 5.2 | The plan only lived in text — a label and a table | Entry / stop / target drawn on the chart, frozen at the exit |
Next — in order
- Plug in the alerts.
alert() with the JSON payload, both legs, trade_id.
Nothing else can start until firings are being recorded.
- Get the filtered sample. Every VWAP pull-in long that also had a coil, a PDH
break and timeframe alignment. That is the number the deck should be quoting, and it doesn't exist yet.
- Settle the hold question. Run the three rules — take 2R · trail the 50-minute line ·
hold to the projected move — over every historical firing. The data picks the rule.
- Then, and only then, put a performance number on a slide as fact.
On deck
AVGO's daily is coiled at its tightest bands of the year, sitting right under the break level — the same
setup loading. If the alert block goes in first, that one gets captured automatically instead of by hand.