SMB CapitalRule Book Evidence Status

Pullback & Reclaim — What We Actually Know

Every rule in the play, with the evidence behind it stated plainly. A rule tested across 4,528 trades and a rule reasoned from one chart are both listed here — labelled so nobody confuses the two. Some of what is on this page contradicts what the checklists said last week; where it does, the test wins.

In Plain English — Start Here

The rest of this page is technical. This part is not. If you read nothing else, read this.

What the trade is. A stock has been going up for a while. It pauses and drifts back down to one of the average-price lines traders watch. It goes quiet there for a few days. Then it closes back above the fastest of those lines — and that is when we buy. We expect it to travel back up to the high it made before the pause.

What we tested. We wrote the rules down precisely, then had a computer apply them to 150 large US stocks over four years — about 4,100 trades. To avoid fooling ourselves, we picked the rules using the first two years only, then scored them on the last two years, which the rules had never seen. If a set of rules only works on the data used to invent it, that shows up immediately.

What we found. The trade works, but the money is made or lost after you are already in. The way we had written the exit rules gave back about half the available profit. A different set of exit rules — same buys, same risk — roughly doubles the result. That is the single finding on this page that matters.

How big is the edge. The average trade makes about 39% of whatever you risked on it, after paying commission and spread — 39 cents back for every dollar risked. That is a real edge but not a large one, and trading costs eat nearly half of it before it reaches you.

The uncomfortable part. More than half of these trades lose money. The profit comes from a minority that run a long way. That means you cannot pick and choose — skipping the one trade that becomes the big winner removes the profit for the month. It also means a run of six or eight losers in a row is completely normal and is not a sign the method has stopped working.

What we still do not know. The four years we tested contain no falling market, so we have no idea how this behaves in one. We tested it only on large, heavily traded stocks. And the computer entered trades in a slightly different way from how the desk actually enters them — close, but not identical. None of those gaps are hidden further down this page; they are listed plainly.

The one-sentence version. The trade is real and worth doing, the exit rules we had written were leaving about half the money on the table, and the evidence is good enough to change those rules but not good enough to bet more money on the strategy.

Evidence Key

Read the tag before you trust the rule.

Tested Backtested across 150 symbols, Oct 2022 – Aug 2026, with a fit/test split so the rule was scored on data that did not choose it.
One trade Derived from CRWD 7/31/2026 alone. Plausible, mechanically sensible, unverified.
Untested Carried over from the original playbook or the desk. Never checked either way.
Open Needs a decision from the desk before it can be written as a rule at all.

Where the Numbers Stand

Best-tested exit ladder, out-of-sample window (2025-01 → 2026-08, n = 1,952).

Mean Rmean RWhat it isThe average outcome per trade, measured in R.What it is forThe headline number for whether a rule makes money. Positive mean R means the average trade wins after wins and losses are netted.How it is derivedAdd every trade's R result and divide by the number of trades.How to read itBelow 0 — do not trade it. 0 to +0.2 — marginal, costs will eat it. +0.2 to +0.5 — a real but modest edge. +0.5 to +1.0 — good. Above +1.0 — very good for a swing strategy, and worth double-checking for data errors. Ours: +0.69 gross, +0.39 after costs — solid, not spectacular.
+0.69
gross, per trade
After costsnet of costsWhat it isMean R after subtracting an estimate of commission and spread.What it is forGross results ignore the cost of getting in and out. A laddered exit pays that cost several times.How it is derivedA 0.15% round trip ÷ the median stop distance (0.91% of price) = 0.16R, charged roughly twice for a laddered exit ≈ 0.33R per trade.How to read itCompare net to gross. If net keeps more than about 70% of gross, the strategy is robust to execution. Below 50%, it is fragile and small slippage changes will break it. Ours: +0.39 of +0.69 = 56% survives — acceptable, not comfortable.
+0.39
0.15% round trip
Median Rmedian RWhat it isThe middle outcome — half of trades did better, half did worse.What it is forShows what a typical trade feels like. Mean and median diverging tells you the result is driven by a few big winners rather than steady gains.How it is derivedSort every trade's R result and take the middle value.How to read itCompare it to the mean. Close together means steady, repeatable gains. Far below the mean means the result depends on a few big winners and most trades disappoint. A median of −1.00 means over half the trades stop out for a full loss. Ours: +0.34 against a +0.69 mean — tail-driven. You must take every signal, because skipping the wrong one costs the year.
+0.34
most trades scratch or lose
95% CI95% CIWhat it isConfidence interval — the range the true average is very likely to sit inside.What it is forReplaces a single point estimate with an honest range. If the range excludes zero, the edge is unlikely to be noise.How it is derivedResample the data 4,000 times and take the 2.5th and 97.5th percentiles of the resulting averages.How to read itLook at two things. Does it include zero? If yes, the edge is not established. And how wide is it? A wide range means genuine uncertainty about the size of the edge. Plan around the lower bound, not the middle. Ours: [+0.207, +1.096] gross — excludes zero, so the edge is real, but after costs the lower end sits at −0.07R.
+0.21 … +1.10
cluster-adjusted, gross
The clustering objection has been tested, and the result survived it. A raw t of 5.65 assumes 1,952 independent bets, which these are not — hundreds fire in the same weeks across names that move together. Resampling whole calendar months with replacement (4,000 draws) keeps correlated trades bundled and gives the honest interval: the out-of-sample 95% CI is [+0.207R, +1.096R] gross, with 0.0% of draws at or below zero. The interval is much wider than the raw t implied, and it does not contain zero before costs. After costs of 0.276R the range becomes [−0.07, +0.82] — the central estimate is clearly positive, the pessimistic end is not.
WindowMonthsMean Rmean RWhat it isThe average outcome per trade, measured in R.What it is forThe headline number for whether a rule makes money. Positive mean R means the average trade wins after wins and losses are netted.How it is derivedAdd every trade's R result and divide by the number of trades.How to read itBelow 0 — do not trade it. 0 to +0.2 — marginal, costs will eat it. +0.2 to +0.5 — a real but modest edge. +0.5 to +1.0 — good. Above +1.0 — very good for a swing strategy, and worth double-checking for data errors. Ours: +0.69 gross, +0.39 after costs — solid, not spectacular.Raw tt-statWhat it isA “could this be luck?” score. How many standard errors the mean sits away from zero.What it is forAbove about 2 the mean is unlikely to be zero by chance; above 3 is strong. It says nothing about how good the strategy is — that is mean R.How it is derivedmean ÷ (standard deviation ÷ √n). Assumes trades are independent, which inflates it here.How to read itUnder 2 — noise, ignore it. 2 to 3 — probably real. 3 to 5 — strong. Above 5 — very strong, but only if trades are independent. Clustered trades make this number look better than it is, which is why the confidence interval below matters more. Ours: 5.65 raw on the uncorrected series — superseded by the confidence interval. 95% CI (month-block)95% CIWhat it isConfidence interval — the range the true average is very likely to sit inside.What it is forReplaces a single point estimate with an honest range. If the range excludes zero, the edge is unlikely to be noise.How it is derivedResample the data 4,000 times and take the 2.5th and 97.5th percentiles of the resulting averages.How to read itLook at two things. Does it include zero? If yes, the edge is not established. And how wide is it? A wide range means genuine uncertainty about the size of the edge. Plan around the lower bound, not the middle. Ours: [+0.207, +1.096] gross — excludes zero, so the edge is real, but after costs the lower end sits at −0.10R.P(mean ≤ 0)P(mean ≤ 0)What it isThe share of bootstrap draws in which the edge disappeared entirely.What it is forA direct answer to “how often would this look profitable and not be?” 0.0% means no resampled version of history lost money.How it is derivedCount the bootstrap draws whose mean was zero or below, divide by 4,000.How to read itUnder 5% — the edge is established. 5 to 10% — borderline, keep testing. Above 10% — not established, do not size up. Ours: 0.0% before costs. After costs the pessimistic tail dips just below zero.
Test 2025-01 → 2026-0820 +0.692— [+0.207, +1.096]0.0%
Full 2022-10 → 2026-0845 +0.770— [+0.430, +1.115]0.0%

Consistency across years, and month-to-month dispersion.

SegmentTradesnWhat it isThe number of trades in the sample.What it is forMore trades means more confidence — but only if they are independent of each other, which clustered trades are not.How it is derivedA straight count of qualifying setups in the window.How to read itUnder 100 — treat as anecdote. 100 to 500 — suggestive. Above 1,000 — solid, if the trades are independent. Clustered trades inflate n: 1,952 trades landing in 20 months is closer to 20 observations than 1,952. Ours: 1,810 tradeable trades across 20 months.Mean Rmean RWhat it isThe average outcome per trade, measured in R.What it is forThe headline number for whether a rule makes money. Positive mean R means the average trade wins after wins and losses are netted.How it is derivedAdd every trade's R result and divide by the number of trades.How to read itBelow 0 — do not trade it. 0 to +0.2 — marginal, costs will eat it. +0.2 to +0.5 — a real but modest edge. +0.5 to +1.0 — good. Above +1.0 — very good for a swing strategy, and worth double-checking for data errors. Ours: +0.69 gross, +0.39 after costs — solid, not spectacular.Medianmedian RWhat it isThe middle outcome — half of trades did better, half did worse.What it is forShows what a typical trade feels like. Mean and median diverging tells you the result is driven by a few big winners rather than steady gains.How it is derivedSort every trade's R result and take the middle value.How to read itCompare it to the mean. Close together means steady, repeatable gains. Far below the mean means the result depends on a few big winners and most trades disappoint. A median of −1.00 means over half the trades stop out for a full loss. Ours: +0.34 against a +0.69 mean — tail-driven. You must take every signal, because skipping the wrong one costs the year.
2022 Q4 — the one drawdown stretch8 +7.40+7.95
2023 — recovery rally1,121+0.953+0.34
2024 — continued rally1,447+1.302+0.58
2025 — out of sample1,169+1.279+0.36
2026 YTD — out of sample783+0.958+0.27

Of 45 months, 12 lose money (27%). No single year carries the result — all four full years land between +0.55R and +1.00R once untradeable stops are excluded.

What the sample still cannot tell you. The 2022 Q4 drawdown segment contains only 8 trades, because the setup requires price above a rising 50-DMA and therefore stops firing in downtrends. That is a feature — the play sits out bear markets rather than losing in them — but it also means there is no bear-market evidence here at all, and none about a chop regime where the 50-DMA rises but trends keep failing. Two further limits stand: the entry tested is the prior-day-high break, not the VWAP pull-in the desk actually used on CRWD 7/31; and stops fill at the exact stop price, with no slippage, in a strategy where over half of trades stop out.

Reading the Scorecard

Each row is one test. Ours is this strategy's result; Bad / Unremarkable / Good are the thresholds that result is judged against; Verdict is where it lands. Hover or tap any metric name for a full explanation of what the row measures and what each threshold means.

MetricOursBadUnremarkableGoodVerdict
Mean R (after costs)Mean R (after costs)What this row measuresThe average profit of a single trade, after commission and spread, measured in risk units. +0.39R means the average trade returns 39 cents for every dollar risked.“Bad” meansBelow zero — the strategy loses money on average. Do not trade it.“Unremarkable” meansBetween zero and +0.3R. It makes money, but so little that a small rise in trading costs or a run of bad fills would wipe it out.“Good” meansAbove +0.5R. Comfortably profitable with room to absorb worse execution than assumed.Where we land+0.39R — between the two. Real and usable, but not a wide margin.+0.39< 00 … +0.3> +0.5Solid — not spectacular
Out-of-sample vs fitOut-of-sample vs fitWhat this row measuresWe chose the rules using 2022–2024 data, then scored them on 2025–2026 data the rules had never seen. This row compares the two scores.“Bad” means“Collapses” — the new-data score falls to near zero or turns negative. The rules were fitted to random noise in the old data and there was never a real effect. This is exactly what happened to a VWAP entry filter we tested and discarded.“Unremarkable” means“Halves” — the new-data score is roughly half the old one. Something real is there, but a good part of the original result was over-fitting.“Good” means“Holds” — the two scores are close. The effect behaves the same way on data it was never shown, which is the strongest evidence available that it is genuine.Where we land+0.692 on new data against +0.770 on the design data — within 10%. It holds.+0.692 vs +0.770collapseshalvesholdsHolds — robust
95% CI excludes zero?95% CI excludes zero?What this row measuresThe confidence interval is the range the true long-run average is very likely to sit inside. This row asks whether that whole range stays above zero.“Bad” means“No” — the range straddles zero, so the data cannot rule out that the strategy has no edge at all.“Unremarkable” means“Barely” — the bottom of the range only just clears zero. A slightly worse assumption anywhere would sink it.“Good” means“Clearly” — the entire range sits well above zero.Where we landYes before costs (+0.21 to +1.10). After costs the bottom dips to −0.10, so net of trading costs it is no longer clear-cut.gross yes / net nonobarelyclearlyYes gross, marginal net
P(mean ≤ 0)P(mean ≤ 0)What this row measuresWe rebuilt history 4,000 times by reshuffling whole months, and counted how often the strategy came out flat or losing. This is the share of those rebuilt histories that failed.“Bad” meansAbove 10% — more than one rebuilt history in ten lost money. Too fragile to rely on.“Unremarkable” means5–10% — borderline. Suggestive but not settled.“Good” meansUnder 5% — the edge shows up in essentially every version of history we can construct.Where we land0.0% before costs. No resampled history lost money gross. After costs the pessimistic tail does dip below zero.0.0%> 10%5–10%< 5%Established (gross)
Costs survivedCosts survivedWhat this row measuresWhat proportion of the gross profit is still there after commission and spread are paid.“Bad” meansUnder 50% — costs eat most of the edge. The strategy is really a bet on execution quality.“Unremarkable” means50–70% — costs are a major drag but the edge survives them.“Good” meansAbove 70% — costs are a manageable overhead rather than the main risk.Where we land56%. Gross +0.69R becomes net +0.39R. Costs take 0.30R from every trade because the exit ladder closes the position in three separate pieces.56%< 50%50–70%> 70%Acceptable, not comfortable
Losing monthsLosing monthsWhat this row measuresThe share of calendar months in which the strategy lost money overall.“Bad” meansAbove 45% — close to a coin flip month to month. Very hard to keep following.“Unremarkable” means35–45% — frequent losing months; expect long uncomfortable stretches.“Good” meansUnder 30% — roughly one losing month in three or better.Where we land27% — 12 of 45 months lost money. Expect roughly one losing month per quarter.27%> 45%35–45%< 30%Good for this style
Median RMedian RWhat this row measuresThe typical trade, as opposed to the average. Half of all trades did better than this and half did worse.“Bad” meansBelow −0.5R — the typical trade is a substantial loser and the whole result depends on rare outliers. Psychologically close to untradeable.“Unremarkable” meansBetween −0.5R and +0.3R — the typical trade roughly breaks even; profits come from the tail.“Good” meansAbove +0.5R — the typical trade wins meaningfully, so results feel steady rather than lumpy.Where we land+0.34R against an average of +0.69R. The typical trade barely scratches; the profit comes from a minority of large winners. Every signal must be taken.+0.34< −0.5−0.5 … +0.3> +0.5Weak — tail-driven
Independent observationsIndependent observationsWhat this row measuresNot the number of trades, but the number of genuinely separate time periods. Hundreds of trades opening in the same week across correlated stocks are effectively one bet, not hundreds.“Bad” meansFewer than 12 — a year or less of distinct conditions. Almost anything could be true.“Unremarkable” means12–36 months — enough to be suggestive, not enough to be confident about the size of the edge.“Good” meansMore than 60 months — five years of varied conditions, including at least one full cycle.Where we land20 months in the out-of-sample window. This is the real constraint on confidence, not the 1,810 trade count.20 months< 1212–36> 60Thin
Bear-market evidenceBear-market evidenceWhat this row measuresHow much of the test period covered a falling market.“Bad” means“None” — the strategy has never been observed in a decline. Its behaviour there is pure conjecture.“Unremarkable” means“One cycle” — one downturn observed. Better than nothing, but a single episode can mislead.“Good” means“Two cycles” — tested through more than one distinct decline, so the behaviour is unlikely to be specific to one event.Where we land8 trades in the only down stretch available. Effectively none. The setup needs a rising 50-day average, so it stops firing in downtrends — it should sit out rather than lose, but that has not been demonstrated.8 tradesnoneone cycletwo cyclesAbsent
Tests the entry actually used?Tests the entry actually used?What this row measuresWhether the backtest triggers trades the same way the desk does in practice.“Bad” means“No” — the test uses a different trigger, so the results describe a different strategy that happens to share the same setup.“Unremarkable” means“Proxy” — a closely related trigger that fires at similar times and prices.“Good” means“Yes” — the exact rule a trader follows.Where we landNo. The test enters on a break of the prior day's high. The desk's CRWD trade was entered on a VWAP pull-in four hours after that break. Resolving which is canonical is the top open item.nonoproxyyesWrong entry
In one line: the statistics are strong, the coverage is thin. Everything measuring whether the tested rule makes money comes back green — the edge survives an honest confidence interval, holds out of sample, and survives costs. Everything measuring whether the tested rule is your rule comes back amber or red: 20 independent months, no bear market, and the wrong entry. The right reading is “a real effect, on a close cousin of our play, in a friendly market” — enough to justify trading it small and finishing the tests, not enough to size up on.
The one number a new reader will misread. Median R of +0.32 against a mean of +1.15 is the important tension on this page. Over half of all trades lose or scratch; the average is positive only because a minority run a long way. Practically that means you cannot cherry-pick signals — skipping the one trade that becomes the outlier removes the year's profit. It also means a run of six or eight losers is normal, not evidence the play has stopped working.

The Rules

Setup and entry.

1 Uptrend, pulled back into a major moving average Untested 5 / 10 / 20 / 50 / 200 day. Name which one. The backtest used a loose proxy for this — the real qualification filters have never been isolated and tested.
2 The base has gone quiet — bands near their tightest in a month Untested BB-width percentile under 20. Carried from the original playbook, never validated. It also cannot separate a "high and tight" coil from a "barcode" — that problem is unsolved.
3 Entry is the break. Do not wait for a pull-back toward VWAP Tested Across 4,518 setups: a 1.5σσ (sigma)What it isStandard deviation — how far price sits from VWAP relative to the day's own scatter.What it is forScales “extended” to the session's volatility, so it means the same thing on a quiet day and a wild one.How it is derivedThe volume-weighted standard deviation of typical price about VWAP, accumulated from the open.How to read itRoughly: under 1σ is normal, 1 to 2σ is extended, over 2σ is stretched. Useful as description. As an entry filter it failed — a 1.5σ band lost money out of sample. VWAPsession VWAPWhat it isVolume-weighted average price — the average price paid so far today, weighted by volume.What it is forA reference for whether buyers or sellers control the session. Above it is constructive; below it is not.How it is derivedΣ(typical price × volume) ÷ Σvolume, accumulated from the opening bell.How to read itAbove it is constructive, below it is not — but do not use it to time an entry. Waiting to buy nearer VWAP tested negative out of sample. entry band returned +2.77R in-sample and −0.47R out of sample. Bands tighter than 2σ lose money. A tighter entry sits closer to its own stop, and the extra stop-outs cost more than the better price earns.
4 Never add a "must be above the 20-DMA" entry filter One trade CRWD entered at 190.86 with the 20-DMA at 192.25. A pullback-to-the-50 setup sits below the fast lines by construction. The mechanism is structural even if the evidence is single.
5 Stop is the hold lowhold lowWhat it isThe lowest low of the entry bar and the two bars before it.What it is forThe structural stop: the level that, if broken, means the setup that produced the trade has failed.How it is derivedmin(low) over the trigger bar and the two preceding bars, on the trigger timeframe.How to read itA structural stop should sit where the idea is wrong, not at a round number or a fixed percentage. If the stop is more than about 2% of price, the setup is probably too extended to take. on the trigger frame Open Two definitions are live and they disagree by 25%: the hold low (187.975 on CRWD) versus the Pine v5.2 literal min(vlLo, lowest(low,3)) (187.18). Every R figure on this page depends on which one is the spec.
6 Size = risk dollars ÷ stop distance Untested The stop sets the size, never the reverse. Median 1RRWhat it isThe risk unit. 1R is the distance from your entry to your stop.What it is forLets trades of different sizes and prices be compared on one scale. A +3R trade made three times what it risked, whatever the share count.How it is derivedentry price − stop price. On CRWD 7/31: 191.21 − 187.98 = 3.23 points, so 1R = 3.23.How to read itNot good or bad by itself — it is the yardstick. What matters is how many R a trade returns. A stop wider than about 2% of price starts to hurt: it buys fewer shares for the same dollars risked. across the sample was 0.91% of price. The per-trade dollar figure is still undecided — the daily loss cap is a separate limit, not a sizing input.

The Exit Ladder — Recommended

This is the part that changed most. The recommendation below beats the ladder currently written into the swing checklist.

30% At 2R2RWhat it isTwice the risk — the first profit target.What it is forWhere the first slice of the position comes off, or where the trailing stop takes over.How it is derivedentry + 2 × (entry − stop). CRWD: 191.21 + 2 × 3.235 = 197.68.How to read itNot good or bad — a level. Worth knowing: taking the whole position off at 2R caps every trade at +2R, which on this data leaves most of the money behind, because the mean is carried by trades that run far past it., start trailtrailWhat it isAn exit that follows price up: you leave on the first close below a chosen line.What it is forKeeps you in a trend without predicting where it ends, and ratchets the exit higher as price rises.How it is derivedCheck the close of each bar against the line on that line's own timeframe. Intraday pokes do not count.How to read itA trail always gives back some of the peak — that is the cost of not predicting the top. Judge it by the average outcome across many trades, never by how much it left on one.ing it — do not sell it Tested Trail on the 5-day-equivalent line (65-period on 30-min). Worth about +0.30R per trade versus selling at 2R, and it held in both windows.
20% Trail the 17-period30-min bar countWhat it isA moving average measured in 30-minute bars rather than days.What it is forLets a trail run faster than a daily line. There are 13 half-hour bars in a session, so 65 bars ≈ 5 trading days.How it is derived65-bar = the average of the last 65 thirty-minute closes. 17-bar ≈ 1.3 days.How to read itMore bars means a slower line and a longer hold. Mean R rises with length out to about 5 days and then flattens — there is no penalty for being slower, only for being too fast. average on the 30-minute chart Tested The fast tranche. Removing it barely changes the mean but drops the median from +0.32 to −0.29 — it buys consistency cheaply.
50% Trail the 17-day moving averageDMAWhat it isDaily moving average — the average close of the last N days.What it is forThe trailing line for the slow tranche. A close below it ends that part of the trade.How it is derived17-DMA = the average of the last 17 daily closes, recalculated each day.How to read itFaster is not better. On this data the 5-DMA was the worst daily trail (+0.698R) and the 17-DMA the best (+1.061R). Faster lines exit on ordinary pullbacks; slower lines stay in trends., daily closes Tested Not the 5-DMA. Out of sample the 5-DMA was the worst daily runner (+0.698R) and the 17-DMA the best (+1.061R). The 17 / 20 / 50-DMA are within noise of each other; the 5-DMA is clearly behind.
LadderFit mean Rfit windowWhat it isThe earlier period used to choose settings. 2022-10 to 2024-12.What it is forAny rule can be tuned to look good on the data used to build it. This window is where that tuning happens — and its results are not trustworthy on their own.How it is derivedA straight date cut at 2025-01-01.How to read itNever judge a rule on its fit-window score. A rule that looks good here and nowhere else was fitted to noise. Treat these numbers as the best case, not the expected case.Test mean Rout of sampleWhat it isThe later period, held back and never used to choose anything. 2025-01 to 2026-08.What it is forThe only honest score. A rule that works in the fit window and fails here was fitted to noise — exactly what happened to the 1.5σ entry band.How it is derivedEvery rule is scored here after being chosen on the fit window.How to read itThis is the number that counts. If out-of-sample is close to fit, the rule is robust. If it collapses, the rule was overfitted. Ours: +0.692 out of sample against +0.770 in fit — close, which is the good outcome.Test medianmedian RWhat it isThe middle outcome — half of trades did better, half did worse.What it is forShows what a typical trade feels like. Mean and median diverging tells you the result is driven by a few big winners rather than steady gains.How it is derivedSort every trade's R result and take the middle value.How to read itCompare it to the mean. Close together means steady, repeatable gains. Far below the mean means the result depends on a few big winners and most trades disappoint. A median of −1.00 means over half the trades stop out for a full loss. Ours: +0.34 against a +0.69 mean — tail-driven. You must take every signal, because skipping the wrong one costs the year.Net @0.15%net of costsWhat it isMean R after subtracting an estimate of commission and spread.What it is forGross results ignore the cost of getting in and out. A laddered exit pays that cost several times.How it is derivedA 0.15% round trip ÷ the median stop distance (0.91% of price) = 0.16R, charged roughly twice for a laddered exit ≈ 0.33R per trade.How to read itCompare net to gross. If net keeps more than about 70% of gross, the strategy is robust to execution. Below 50%, it is fragile and small slippage changes will break it. Ours: +0.39 of +0.69 = 56% survives — acceptable, not comfortable.
A — 30 @2R / 50 @17bar / 20 @5-DMA (current checklist) +0.674+0.519+0.94+0.216
D — 30 trailed / 20 / 50 @5-DMA +0.758+0.598+0.46+0.294
D17 — 30 trailed / 20 / 50 @17-DMA (recommended) +0.830+0.692+0.34 +0.389
D20 — …50 @20-DMA +0.795+0.675+0.33+0.372
D50 — …50 @50-DMA +0.951+0.736+0.25+0.432
30 @2R / 70 @5-DMA (two tranche) +0.530+0.450−0.49+0.147
D50 edges D17 on this corrected data — treat them as tied. Once untradeable stops are excluded, the 50-DMA runner scores marginally higher (+0.736 against +0.692 out of sample). The gap is well inside the noise of a 20-month sample. What is robust across every cut is the shape: trail the 30% rather than selling it, and run the slow tranche on the 17-DMA or slower — never the 5-DMA.
The trade-off to understand before adopting D17. Ladder A wins more often — median +0.94 against D17's +0.32. D17 makes more on average because its tail is fatter. You are trading a higher hit rate for a bigger average outcome. That is a temperament decision as much as a mathematical one.
More than half of these trades lose. Median Rmedian RWhat it isThe middle outcome — half of trades did better, half did worse.What it is forShows what a typical trade feels like. Mean and median diverging tells you the result is driven by a few big winners rather than steady gains.How it is derivedSort every trade's R result and take the middle value.How to read itCompare it to the mean. Close together means steady, repeatable gains. Far below the mean means the result depends on a few big winners and most trades disappoint. A median of −1.00 means over half the trades stop out for a full loss. Ours: +0.34 against a +0.69 mean — tail-driven. You must take every signal, because skipping the wrong one costs the year. for any all-out variant is −1.00: over half stop out for a full loss, and the positive mean comes entirely from a fat right tail. The banker tranches are what pull the median back above zero. This only works at consistent size over many trades — which is the same conclusion the CRWD review reached, where the position made money and the day-session round trips lost it.

Management & Exit Scope

7 The 3:50 check runs on the entry day only One trade Above the resistance necklinenecklineWhat it isThe resistance level that capped the base before the break.What it is forUsed by the 3:50 check as the line price must hold above to justify carrying the trade overnight.How it is derivedThe horizontal resistance the coil broke through. CRWD 7/31: 189.62.How to read itA static level, so it stops being informative once price runs well above it. That is why the 3:50 check that uses it is an entry-day rule, not a daily one. and above session VWAPsession VWAPWhat it isVolume-weighted average price — the average price paid so far today, weighted by volume.What it is forA reference for whether buyers or sellers control the session. Above it is constructive; below it is not.How it is derivedΣ(typical price × volume) ÷ Σvolume, accumulated from the opening bell.How to read itAbove it is constructive, below it is not — but do not use it to time an entry. Waiting to buy nearer VWAP tested negative out of sample. at 15:50 → carry overnight. Either fails → flatten. Run daily it flattens CRWD on 8/05 at +6.35R and never sees 225.53: the neckline is static, so after day one only the VWAP test binds.
8 A trail means a close below the line, on that line's own timeframe Untested Intraday pokes do not count. This is a definition, not a claim — but note it was assumed throughout the backtest, so the results depend on it.
9 There is no 2–4 day "dead zone" in trail length Tested Mean R rises monotonically with trail length out to about 5 days, then flattens: 13-bar +1.02 → 26-bar +1.13 → 34-bar +1.29 → 65-bar +1.33 → 91-bar +1.32. An earlier claim that 2–4 days underperformed was an artifact of one CRWD consolidation.

Corrections Log

Ideas that looked right and were not. Recorded so they are not re-proposed from first principles.

“+1.15R per trade” — the headline number itself Inflated by backtest trades whose stop sat a fraction of a percent from entry: one showed +3,872R on a 0.002% stop. Untradeable in practice, but they supplied half of all recorded R. Excluding stops under 0.25% of price cuts the result to +0.69R gross, +0.39R net — the edge survives at roughly half the size.
“The VWAP pull-in entry nearly doubles the return” Same artifact, worse: the pull-in enters near VWAP and produces tight stops far more often, so 20% of its trades were untradeable against 7% for the break entry. Corrected, the break entry is better — 0.451% of price net against 0.297%.
“Enter within 1.5σ of VWAP” Best variant in-sample at +2.77R, lost money out of sample at −0.47R. The textbook overfit — and it would have looked validated if only the fit window had been run.
“The 5-DMA is the best trail line” True on CRWD. Across 1,952 out-of-sample trades it is the worst of the daily runners — +0.698R against the 17-DMA's +1.061R.
“Flip the ladder weights to 30/20/50” The current 30/50/20 beat it in both windows. The reasoning — put size on the slow tranche — sounded structural and was wrong.
“Two tranches are simpler and better” The worst ladder tested, in both windows. It won on CRWD alone.
“A 2–4 day dead zone in trail length” Not real. It was one CRWD consolidation viewed three ways and mistakenly described as three observations.
“225.53 was the objective” That was the peak the trade made. The objective — the prior swing high — was 217.50, set on 7/15, before the trade existed.
“Entry is the first break of the prior-day high” CRWD's prior-day high was broken at the opening bell on 7/31; the fill came four hours later. The entry actually taken was a VWAP pull-in. The whole backtest above tests the break entry, not the one the desk used.

What Would Make This a Real Edge

In priority order. Each is a specific, runnable check.

✓ Block bootstrap the significance Done Month-block resampling, 4,000 draws. Out-of-sample CI [+0.207, +1.096] gross, 0.0% of draws at or below zero. The clustering objection was real — the interval is far wider than the raw t suggested — but the edge survived it. After costs the range is [−0.10, +0.79].
✓ Segment by regime Done All four full years land between +0.55R and +1.00R — no single year carries it. But the only drawdown stretch holds just 8 trades, so bear-market behaviour remains unknown.
1 Re-run on the entry the desk actually uses The VWAP pull-in, with its ride / leg / rail arming conditions — not the prior-day-high break. This is currently blocked: the engine's LEG filter misses by 1.6% of ATR and will not fire the reference trade unaided.
4 Add the real qualification filters BB-width percentile, coil bar-count, volume pocket, MA cluster width. The backtest's setup filter is a loose proxy; the desk's is tighter and may be doing most of the work — or none of it.
5 Model stop slippage Stops currently fill at the exact stop price. Over half these trades stop out, so this matters more here than in most strategies.
6 Get the loser set out of the pocket-pivot scanner The 18-setup example library is entirely trades that worked. No qualification filter can be validated without the ones that failed — ma_state.py is already built for exactly this and needs winners and losers at n ≥ 30 per bucket.
7 Test a small-cap universe All of the above is 150 liquid large caps. ALOY, BTDR and BBIO are not in it and do not behave like it.

Open Decisions — Desk Only

These cannot be settled by testing. They need a ruling.

DecisionOptionsWhat it changes
Stop conventionhold low (3.235) / Pine literal (4.030) every R figure, by 25%
Canonical entryPDH break / VWAP pull-in what the backtest should be testing at all
1R overnight gateon / off on CRWD it would have flattened at +0.11R and missed the move
Per-trade risk $—share sizing; the daily loss cap is separate
Earnings policyhold / flat into the printany runner spanning a report

Glossary

Every metric on this page, in full. The same text appears as a tooltip wherever the term is used — hover it, or tap it on a phone.

RWhat it isThe risk unit. 1R is the distance from your entry to your stop.What it is forLets trades of different sizes and prices be compared on one scale. A +3R trade made three times what it risked, whatever the share count.How it is derivedentry price − stop price. On CRWD 7/31: 191.21 − 187.98 = 3.23 points, so 1R = 3.23.How to read itNot good or bad by itself — it is the yardstick. What matters is how many R a trade returns. A stop wider than about 2% of price starts to hurt: it buys fewer shares for the same dollars risked.
mean RWhat it isThe average outcome per trade, measured in R.What it is forThe headline number for whether a rule makes money. Positive mean R means the average trade wins after wins and losses are netted.How it is derivedAdd every trade's R result and divide by the number of trades.How to read itBelow 0 — do not trade it. 0 to +0.2 — marginal, costs will eat it. +0.2 to +0.5 — a real but modest edge. +0.5 to +1.0 — good. Above +1.0 — very good for a swing strategy, and worth double-checking for data errors. Ours: +0.69 gross, +0.39 after costs — solid, not spectacular.
median RWhat it isThe middle outcome — half of trades did better, half did worse.What it is forShows what a typical trade feels like. Mean and median diverging tells you the result is driven by a few big winners rather than steady gains.How it is derivedSort every trade's R result and take the middle value.How to read itCompare it to the mean. Close together means steady, repeatable gains. Far below the mean means the result depends on a few big winners and most trades disappoint. A median of −1.00 means over half the trades stop out for a full loss. Ours: +0.34 against a +0.69 mean — tail-driven. You must take every signal, because skipping the wrong one costs the year.
nWhat it isThe number of trades in the sample.What it is forMore trades means more confidence — but only if they are independent of each other, which clustered trades are not.How it is derivedA straight count of qualifying setups in the window.How to read itUnder 100 — treat as anecdote. 100 to 500 — suggestive. Above 1,000 — solid, if the trades are independent. Clustered trades inflate n: 1,952 trades landing in 20 months is closer to 20 observations than 1,952. Ours: 1,810 tradeable trades across 20 months.
t-statWhat it isA “could this be luck?” score. How many standard errors the mean sits away from zero.What it is forAbove about 2 the mean is unlikely to be zero by chance; above 3 is strong. It says nothing about how good the strategy is — that is mean R.How it is derivedmean ÷ (standard deviation ÷ √n). Assumes trades are independent, which inflates it here.How to read itUnder 2 — noise, ignore it. 2 to 3 — probably real. 3 to 5 — strong. Above 5 — very strong, but only if trades are independent. Clustered trades make this number look better than it is, which is why the confidence interval below matters more. Ours: 5.65 raw on the uncorrected series — superseded by the confidence interval.
95% CIWhat it isConfidence interval — the range the true average is very likely to sit inside.What it is forReplaces a single point estimate with an honest range. If the range excludes zero, the edge is unlikely to be noise.How it is derivedResample the data 4,000 times and take the 2.5th and 97.5th percentiles of the resulting averages.How to read itLook at two things. Does it include zero? If yes, the edge is not established. And how wide is it? A wide range means genuine uncertainty about the size of the edge. Plan around the lower bound, not the middle. Ours: [+0.207, +1.096] gross — excludes zero, so the edge is real, but after costs the lower end sits at −0.10R.
block bootstrapWhat it isResampling whole calendar months instead of individual trades.What it is forTrades cluster — hundreds fire in the same week across names that move together. Resampling by month keeps correlated trades bundled so the confidence interval is not falsely narrow.How it is derivedDraw 45 months at random with replacement, pool their trades, take the mean. Repeat 4,000 times.How to read itThere is no good or bad value here — it is a method, not a metric. Its job is to stop you fooling yourself. If a result survives block bootstrapping it is probably not an artefact of one lucky stretch of market.
P(mean ≤ 0)What it isThe share of bootstrap draws in which the edge disappeared entirely.What it is forA direct answer to “how often would this look profitable and not be?” 0.0% means no resampled version of history lost money.How it is derivedCount the bootstrap draws whose mean was zero or below, divide by 4,000.How to read itUnder 5% — the edge is established. 5 to 10% — borderline, keep testing. Above 10% — not established, do not size up. Ours: 0.0% before costs. After costs the pessimistic tail dips just below zero.
fit windowWhat it isThe earlier period used to choose settings. 2022-10 to 2024-12.What it is forAny rule can be tuned to look good on the data used to build it. This window is where that tuning happens — and its results are not trustworthy on their own.How it is derivedA straight date cut at 2025-01-01.How to read itNever judge a rule on its fit-window score. A rule that looks good here and nowhere else was fitted to noise. Treat these numbers as the best case, not the expected case.
out of sampleWhat it isThe later period, held back and never used to choose anything. 2025-01 to 2026-08.What it is forThe only honest score. A rule that works in the fit window and fails here was fitted to noise — exactly what happened to the 1.5σ entry band.How it is derivedEvery rule is scored here after being chosen on the fit window.How to read itThis is the number that counts. If out-of-sample is close to fit, the rule is robust. If it collapses, the rule was overfitted. Ours: +0.692 out of sample against +0.830 in fit — close, which is the good outcome.
net of costsWhat it isMean R after subtracting an estimate of commission and spread.What it is forGross results ignore the cost of getting in and out. A laddered exit pays that cost several times.How it is derivedA 0.15% round trip ÷ the median stop distance (0.91% of price) = 0.16R, charged roughly twice for a laddered exit ≈ 0.33R per trade.How to read itCompare net to gross. If net keeps more than about 70% of gross, the strategy is robust to execution. Below 50%, it is fragile and small slippage changes will break it. Ours: +0.39 of +0.69 = 56% survives — acceptable, not comfortable.
2RWhat it isTwice the risk — the first profit target.What it is forWhere the first slice of the position comes off, or where the trailing stop takes over.How it is derivedentry + 2 × (entry − stop). CRWD: 191.21 + 2 × 3.235 = 197.68.How to read itNot good or bad — a level. Worth knowing: taking the whole position off at 2R caps every trade at +2R, which on this data leaves most of the money behind, because the mean is carried by trades that run far past it.
DMAWhat it isDaily moving average — the average close of the last N days.What it is forThe trailing line for the slow tranche. A close below it ends that part of the trade.How it is derived17-DMA = the average of the last 17 daily closes, recalculated each day.How to read itFaster is not better. On this data the 5-DMA was the worst daily trail (+0.698R) and the 17-DMA the best (+1.061R). Faster lines exit on ordinary pullbacks; slower lines stay in trends.
30-min bar countWhat it isA moving average measured in 30-minute bars rather than days.What it is forLets a trail run faster than a daily line. There are 13 half-hour bars in a session, so 65 bars ≈ 5 trading days.How it is derived65-bar = the average of the last 65 thirty-minute closes. 17-bar ≈ 1.3 days.How to read itMore bars means a slower line and a longer hold. Mean R rises with length out to about 5 days and then flattens — there is no penalty for being slower, only for being too fast.
trailWhat it isAn exit that follows price up: you leave on the first close below a chosen line.What it is forKeeps you in a trend without predicting where it ends, and ratchets the exit higher as price rises.How it is derivedCheck the close of each bar against the line on that line's own timeframe. Intraday pokes do not count.How to read itA trail always gives back some of the peak — that is the cost of not predicting the top. Judge it by the average outcome across many trades, never by how much it left on one.
trancheWhat it isA slice of the position managed under its own rule.What it is forLets one trade both bank a certain profit and stay in for a large move, instead of choosing between them.How it is derivedThe ladder here is 30% / 20% / 50% of the original share count.How to read itMore tranches is not automatically better. Two tranches tested worst here; three tested best. The value comes from each slice having a genuinely different job.
hold lowWhat it isThe lowest low of the entry bar and the two bars before it.What it is forThe structural stop: the level that, if broken, means the setup that produced the trade has failed.How it is derivedmin(low) over the trigger bar and the two preceding bars, on the trigger timeframe.How to read itA structural stop should sit where the idea is wrong, not at a round number or a fixed percentage. If the stop is more than about 2% of price, the setup is probably too extended to take.
session VWAPWhat it isVolume-weighted average price — the average price paid so far today, weighted by volume.What it is forA reference for whether buyers or sellers control the session. Above it is constructive; below it is not.How it is derivedΣ(typical price × volume) ÷ Σvolume, accumulated from the opening bell.How to read itAbove it is constructive, below it is not — but do not use it to time an entry. Waiting to buy nearer VWAP tested negative out of sample.
σ (sigma)What it isStandard deviation — how far price sits from VWAP relative to the day's own scatter.What it is forScales “extended” to the session's volatility, so it means the same thing on a quiet day and a wild one.How it is derivedThe volume-weighted standard deviation of typical price about VWAP, accumulated from the open.How to read itRoughly: under 1σ is normal, 1 to 2σ is extended, over 2σ is stretched. Useful as description. As an entry filter it failed — a 1.5σ band lost money out of sample.
prior-day highWhat it isThe highest price of the previous session.What it is forThe breakout level: taking it out signals the coil has resolved upward.How it is derivedmax(high) over the previous trading day.How to read itJust a level. Its value is that it is unambiguous and knowable in advance, which makes it testable — unlike 'looks like it's breaking out'.
reclaimWhat it isA close back above a moving average that price had been trading below.What it is forThe entry signal. It is evidence sellers are exhausted — distinct from a touch, which is price merely reaching the line and turning.How it is derivedRequires all three: price was below, price is now above, and it is a close, not a wick.How to read itThe whole edge of the setup sits in waiting for the close. Buying the touch instead is the single most expensive mistake available here.
necklineWhat it isThe resistance level that capped the base before the break.What it is forUsed by the 3:50 check as the line price must hold above to justify carrying the trade overnight.How it is derivedThe horizontal resistance the coil broke through. CRWD 7/31: 189.62.How to read itA static level, so it stops being informative once price runs well above it. That is why the 3:50 check that uses it is an entry-day rule, not a daily one.
objectiveWhat it isThe prior swing high — the level price is travelling back toward.What it is forThe natural profit target. On the weekly timeframe it replaces a fixed 2R as the first scale-out.How it is derivedThe highest high of the advance that preceded the pullback. CRWD: 217.50, set 7/15.How to read itA target, not a stop. Reaching it is where you scale, not necessarily where you exit — CRWD ran 8 points past its objective before turning.