ICT Says Half the Opening Gap Fills by 10 A.M. 70% of the Time. We Tested the Wrong Side Too.

The one number ICT gives

Late in Part 1, on a live Nasdaq chart a few minutes before the 09:30 open, ICT gives new students a number:

He repeats it in Part 2 (56:51) and Part 4 (16:07). Across five videos that run to about five hours and forty minutes, it is the only hit rate he gives for a market event.

He does not oversell it. When he gives the figure in Part 4, his next words are “Don’t hold me to that number.” (16:09). Then he asks students to keep a record every day and “at the end of the year do a number on the statistical … probability of that having happened” (16:21). The count, he says, will show a recurring phenomenon that students could take advantage of (16:28). In Part 1 he goes further: students should measure the statistical probabilities of what he teaches for themselves, and either see the evidence or watch the ideas fall apart (P1 1:13:17). That is why counting it is fair.

So we did the count, on ten years of NQ futures minute data instead of one year, and added the one thing a daily journal never records: a comparison level the same distance from the open, in the opposite direction.

The last “70%” we chased in this series was a win rate advertised for ICT-style session strategies. This one is ICT’s own, and it is a hit rate for a price level.

What ICT teaches in “How New Students Should Start”

The series is five videos on ICT’s YouTube channel, published in August and September 2026 for people who are new to his material.

Times are (h:)mm:ss into each video; “P4 04:21” means Part 4 at 4:21.

The three pillars

Part 4 is the core. ICT gives beginners three things to log every trading day on the Nasdaq, which he calls the three pillars to start with (P4 04:21):

1. The pre-market, 07:00–09:00 ET (P4 05:43): its high and low, its relative equal highs and lows, and whether it trended or consolidated.
2. The regular-hours opening range gap (ORG): the distance from the prior session’s 16:14 ET one-minute close to the 09:30 open, and whether its midpoint, the CE, trades by 10:00.
3. The Silver Bullet: the first fair value gap after 10:00 ET (P4 18:44). A fair value gap is a three-candle pattern in which the first and third candles do not overlap, leaving a band of price that only the middle candle traded through.

The regimen is modest. He asks students to pick one pillar for the first one or two weeks, add a second, and add the third by the fifth and sixth week, then log them every day (P4 02:25–02:53), about 20 minutes a day (P4 26:26), for 12 weeks of data on one market, the Nasdaq (P4 25:03–25:28).

What the series is, and is not

It is journaling and paper practice, with no stated performance. ICT is plain about that:

In Part 2 he says a brand-new student would not be expected to take a trade, not even on a demo account (P2 1:06:41). In Part 3 he says he has never promised students profitability (P3 29:02). He hedges the pillars too: the pre-market read, he says, is not a panacea and will not work all the time (P4 10:42).

He does say the log will prove to students that these things work and repeat (P4 25:59). So the fair question is whether these events recur more often than chance would give. We do not test whether journaling makes someone a better trader.

It would also be wrong to say ICT told students not to backtest. Part 5 is titled “Back Testing Properly”. There he describes backtesting as logging a hypothetical entry and a hypothetical stop against how price actually delivered (P5 21:59), and as going back to see how many times these things work (P5 05:18).

He also anticipates articles like this one. Of critics who say they ran the numbers on his material, he says:

In context, the rules are his classification of the first presented fair value gap, taught in another video that week. That is a fair point, and we concede it. We coded the rules stated in these five videos as closely as we could, and where he gives no trade rule we used the simplest mechanical proxy. Where he gives two versions (a simple 10:00 fair value gap recipe and his Part 5 walkthrough), we tested both. The discretionary layer was not tested (see the limitations), and nothing here reaches a discretionary trader.

Why levels look right in hindsight, and the comparisons we used

ICT is open that the log is kept in hindsight: at first a student looks for the pattern where it is already obvious in the past (P5 11:42), and he asks students to write their annotations confidently and positively (P5 11:48–12:00). That suits training the eye rather than keeping score. The trap is that most levels within reach of the open get touched on most mornings, so a folder of touched levels is not evidence on its own. Each claim below gets a comparison.

Comparison 1, for levels: the mirror. It sits the same distance from the 09:30 open as the CE, but on the side away from the prior close. It shares the day, the volatility and the distance; only the gap’s direction differs. Say NQ opens 100 points below the prior close. The CE is 50 points above the open, and the mirror is 50 points below it. If the gap pulls price, the CE has to beat its mirror clearly.

Comparison 2, for trades: the same order on a random day. Same order type, same placement minute, and a volatility-scaled stop and target, drawn 1,000 times. The Part 5 recipe uses dates within ±30 sessions and 200 replicates, keeps the real stop and target distances, and is tested both as a market entry and as a like-for-like resting limit. Our placebo test of four ICT setups explains why the random baseline, not zero, is the bar to beat.

Comparison 3, for many claims at once. One fixed statistic per claim, with a Holm correction across ten claims, so this article cannot pick the best of the hundreds of cells the analyses print. That family was declared on September 28, 2026, after the analyses were run, so it is not a pre-registration.

Data. Databento NQ one-minute futures bars, January 4, 2016 to September 4, 2026, with ICT’s 16:14 ET prior close. Replications use Nasdaq, S&P 500 and Dow CFDs and 508 US large caps; the method note at the end has the details.

The count: consequent encroachment by 10:00

ICT sets no minimum gap, so we counted every non-zero gap. A level counts as touched if the running high or low of the 09:30–09:59 bars reaches it. The 42 contract-roll days and 6 days with no gap are skipped.

The CE traded by 10:00 on 1,522 of 2,636 sessions: 57.7% [55.8, 59.6]. Against a literal 70%, the exact binomial p is 2.2 × 10⁻⁴⁰. That test sits outside the ten-claim family, and it would survive any correction. On gaps of at least 2 points (the filter the CFD comparison below uses), the answer is the same: 56.6% (n 2,554). The CFD versions of NQ, the S&P 500 and the Dow come in lower, at 49.6–54.3% (table below).

Still, 58% is most mornings. The question is whether the gap causes it.

The same distance in the opposite direction gets hit almost as often

The mirror level was touched by 10:00 on 56.1% [54.2, 58.0] of the same days. CE minus mirror is +1.6 pp [−1.0, +4.3]; a month-block bootstrap gives [−1.2, +4.5]. The CE was hit alone on 651 days and the mirror alone on 608, for an exact McNemar p of 0.24.

Put another way, the mirror reproduces 97% of the CE’s rate (56.1 ÷ 57.7), and at least 92% even at the upper end of the confidence interval. The data rule out a CE advantage of more than about 4 percentage points.

The chart shows the whole curve, from one-eighth of the gap to a full fill. On both sides the touch rate falls smoothly with distance.

A second check points the same way. Against levels at the same volatility-scaled distance on random dates, the CE sits between the 74th and the 95th percentile across four placebo designs (1,000 draws each). None is significant at 5%, two-sided.

The CFDs are Dukascopy bid one-minute bars, with gaps of at least 2 NQ-point equivalents. None of the five differences is significant. On the NQ and Dow CFDs the difference is about zero, and the S&P CFD’s +3.1 pp (p 0.055) is fragile: a small tilt toward the gap at every distance, not a CE effect.

Where “about 70%” comes from: gap size

Gap size here is measured in percent of the prior close, not in points. NQ’s price rose several-fold over the decade, so a fixed number of points meant a much bigger move in 2016 than in 2026.

On the smallest gaps, under 0.10% of the prior close, the CE is touched by 10:00 on 93.6% of days and the mirror on 92.9% (n 408). On gaps of 1% or more it is 17.5% against 14.9% (n 388).

Cutting the sample at gaps under 0.5% of price (n 1,629) gives 74.0% for the CE and 72.6% for the mirror: +1.4 pp [−1.9, +4.7], p 0.42. That is where “about 70%” comes from. On small gaps, ordinary opening movement covers half the gap, and it covers the same distance in the opposite direction almost as often.

Finer 0.1%-wide bands swing both ways. Gaps of 0.2–0.3% show the CE ahead by +10.9 pp [+2.3, +19.2] (n 322); gaps of 0.4–0.5% show it behind by −12.7 pp [−22.3, −2.8] (n 260). Neither survives a correction for scanning the 11 bands (family-wise p 0.12); in a finer scan of 57 overlapping 0.1% windows, each band’s adjusted p is about 0.40. So the mirror does not match the CE band for band, but no band reliably favours the CE either.

Where the midpoint does edge ahead

The data do not say the CE can never matter. Two samples show it ahead of its mirror.

Earlier NQ data, 2010–2015. On a 15:59 reference, counting days with all 30 opening bars (n 628), CE minus mirror is +7.5 pp [+2.2, +12.7], p 0.0067. The same reference from 2016 gives +0.9 pp. The difference between the eras has p 0.03, but that comparison is post hoc, and coverage before 2016 is sparse.

US stocks. Across 508 US large caps and 1,166,254 stock-days (five-minute bars, 15:55 prior close, gaps between 0.05% and 15%), the CE is hit 72.4% of the time and the mirror 69.5%: +2.98 pp [+2.31, +3.66], or +2.70 pp [+2.02, +3.38] after removing stale prior-close prints. The asymmetry peaks at 30% of the gap (+3.25 pp), not at the CE. It is a small, generic fade, not a midpoint magnet, and a universe of today’s large caps is survivor-biased.

So a small, generic tendency to fade gaps shows up in some samples. The claim that price trades to the midpoint 70% of the time because it is the midpoint does not show up on NQ since 2016. (Consequent encroachment is used differently in our fair value gap backtest, as the midpoint of a fair value gap.)

The three daily pillars as mechanical trades

Part 5 walks through a hypothetical entry, stop and target for two examples: a fair value gap qualified by the pre-market high, and the Silver Bullet. For the pre-market sweep and the opening-gap fade he gives targets but no complete entry and exit rules, and he gives no statistics for any of the pillars as trades. His log is kept in hindsight. To ask whether the levels carry an edge, we used a simple mechanical trade for each pillar and asked three questions:

1. Is the average gross R different from zero?
2. Does it beat the same order placed on random days?
3. What happens net of an assumed 1 NQ point per round turn?

None of the four versions shows a detectable gross edge, and none beats like-for-like random entries. After a 1-point cost, every version loses on average; whether that loss is statistically significant depends on the cost and the tie rule. No tie rule or fill model makes the pre-market sweep, the gap fade or recipe A significant after the ten-claim correction (the most extreme cell reaches Holm 0.079).

Pre-market sweep: resting limit, touch fill, flat at 15:59. ORG fade: market order at the 09:30 open, target fills on touch, flat at the 09:59 close. Silver Bullet recipe A: limit at the near edge, touch fill, flat at 15:59. Recipe B (Part 5): limit 1 tick short of the CE, filled only when the CE trades, flat at 16:00. Stops fill at their level with no slippage. Net figures subtract 1 NQ point per round turn, an assumption. Random-entry percentiles use 1,000 draws (Part 5: ±30 sessions, 200 replicates).

The same trades on other markets tell a similar story. On the S&P 500 and Dow CFDs, none of the three simple trades has a gross edge that survives a cost of 1 NQ-point equivalent. The one near-miss is the S&P opening-gap fade (gaps of at least 0.1%, market fill at the 09:30 open, flat at 09:59): gross +0.039 / +0.044 / +0.049R (stop / neutral / target; p 0.063 / 0.037 / 0.021), and net of 1 NQ-point equivalent −0.012 / −0.007 / −0.002R.

Pillar 1: the pre-market range, 07:00–09:00 ET

ICT asks students to mark the pre-market range and its relative equal highs and lows (P4 05:43). Price sweeping below the pre-market’s relative equal lows and then aiming for the other end of the 07:00–09:00 range is, he says, a “bread and butter” setup (P4 10:53–11:15). We code the simpler version, a sweep of the range extreme. And a trending pre-market tends to be followed by consolidation after 09:30, and vice versa (P4 09:00–10:20; of an overnight trend, P2 29:35). He calls the pre-market “the proverbial crystal ball”, and in the same breath says it is not a panacea (P4 10:25–10:45).

The trade. The pre-market range must span at least 100 one-minute bars and at least 5 points, with the 09:30 open inside it. Sweeps are searched from 09:30. A resting limit order sits at the swept level (touch fill), the target is the other side of the range, and the stop is one range beyond (1:1). The trade goes flat at 15:59, and the result is the same under all tie rules.

Result. Gross −0.045R [−0.088, −0.002] over 1,954 trades, raw p 0.040. But the adjusted p across the sensitivity grid is 0.37, the ten-claim Holm p is 0.36, and against random entries it sits at the 25th percentile (p 0.50). No edge detected before costs; net of 1 point it is −0.079R.

The definition matters. Searching for sweeps from 09:00 instead of 09:30 gives +0.002R gross (n 2,640, touch fill, neutral path). Relative-equal pools are not coded.

The phase claim. To measure how trending the pre-market was, we used an efficiency ratio (ER5): the net move from 07:00 to 08:59 divided by the sum of the absolute five-minute moves, so 1 is a straight trend and near 0 is chop. Its rank correlation with the 09:30–10:59 range (scaled by the trailing 20-day median) is +0.037 [−0.001, +0.076], p 0.165 (n 2,678). ICT predicts a negative sign. The observed sign is the opposite, and it is not significant. Controlling for the size of the pre-market range gives an ICT-shaped partial correlation of −0.154, but random-sign paths built from the same bars give −0.161 (p 0.68): an artefact of the measure.

The takeaway: pre-market volatility forecasts post-open volatility, and the pre-market’s shape adds nothing. For a related pre-open range test, see the New York 8:30 range reversal in our session-strategies backtest.

Pillar 2: fading the opening gap toward its midpoint

ICT suggests anchoring a target to the CE when a student is already looking for prices in that direction, paired with an entry of the student’s choosing once they have learned his entry patterns (P4 16:34). Our ORG fade is the simplest mechanical proxy for that, and it is our simplification, not his rule: it drops that directional condition and takes every qualifying gap.

The trade. Gaps of at least 10 points; a market order at the 09:30 open toward the prior close; the target is the CE, with a 1:1 stop; the target fills on touch; flat at the 09:59 close. The 10-point filter matters because on smaller gaps the brackets can be as narrow as 1 point, and the tie rule alone decides the sign (stop-first −0.061R, target-first +0.090R on gaps of 2 points or more).

n 2,188 trades, gaps ≥ 10 pt. Break-even cost is the round-turn cost in NQ points at which the average trade nets zero.

Gross, the fade is flat to slightly positive depending on the tie rule. Against random entries (gross, stop-first, daily-volatility scaling) its edge is +0.019R [−0.019, +0.057]: the 84th percentile, p 0.32, and between the 71st and the 90th across 13 combinations of volatility scaling and tie rule. Net of 1 point it is negative under every tie rule, but that loss is significant after correcting for the 144-cell sensitivity grid only under stop-first ties (adjusted p 0.022). A “realistic” cost is itself an assumption: fees plus one tick per market, stop or time leg come to 0.60–0.62 pt on NQ and 1.00–1.02 pt on MNQ. The tie rule decides 2.8% of these trades.

Pillar 3: the Silver Bullet, backtested two ways

ICT’s definition for new students is the first fair value gap after 10:00 ET (P4 18:44), and in Part 5 he stresses it is the very first one (P5 33:00). Part 5 also narrates a full walkthrough, so we coded two recipes. Our earlier Silver Bullet study found that the full model’s direction depends on a preselected draw on liquidity; the new-student version drops that choice. For a neighbouring idea, see our first-hour fair value gap test.

Recipe A, simple. The first one-minute fair value gap whose second candle is at or after 10:00 and whose third candle is at or before 10:59. A limit order rests at the gap’s near edge until 11:00 (touch fill). The stop is 1 tick beyond candles 1–3 (risk at least 2 points), the target 1 tick beyond the extreme printed between formation and fill, and the trade goes flat at 15:59.

Gross, recipe A ranges from −0.033R [−0.077, +0.011] under stop-first ties to +0.017 under target-first, with the 50/50 midpoint at −0.008 (n 2,344). The neutral path (−0.026) is adverse to resting limit orders, so the band and the midpoint are the fairer summary. A stricter fill that needs price to trade 1 tick through the entry moves every figure lower (table).

Gross average R per trade unless marked net; net subtracts 1 NQ point per round turn.

Against random entries recipe A sits at the 26th percentile (p 0.52), and between the 2nd and the 27th across 13 combinations of volatility scaling and tie rule. Only the lowest cell (2nd percentile, nominal p 0.04: daily scaling with target-first ties) dips below 5%, and it does not survive local-volatility scaling, where recipe A sits at the 5th to 10th percentile and no cell is significant. Against the opposite trade in the same gap, on the 94.5% of days that one-minute bars fully resolve, the difference is +0.054R [−0.027, +0.136], so the direction is not established either way.

Recipe B, as narrated in Part 5. The first fair value gap of 10:00–10:59 in either direction. Volume imbalances (body gaps between its candles) extend the zone, and the order is armed only once the 10:00 hour has printed an extreme beyond the gap’s candles on the target side (his lunch-macro condition). A limit order sits 1 tick short of the CE and fills only when the CE trades. The stop goes 1 tick beyond the last candle that opened outside the zone. Half the position exits 1 tick before the 10:00-hour extreme, the runner 1 tick before the pre-market extreme. Our additions, not his: a breakeven stop after the first target, risk of at least 2 points, fills before 11:00, and flat at 16:00.

Gross, recipe B makes −0.003R [−0.098, +0.092] under the neutral path (−0.005 stop-first, +0.025 target-first), or −0.09 points per contract [−1.38, +1.20], over 1,454 trades.

n 1,454, neutral path, the Part 5 recipe as specified above.

The loss in R becomes significant only at 0.75 points or more per round turn, and per contract it is not significant at any cost (−1.09 pt [−2.38, +0.20] at 1 point). The reason is that a fixed cost in points weighs more in R when the risk is small: for this trade one tick is 0.08R in 2016 and 0.01R in 2026.

Against random entries on nearby dates, recipe B gives a market-order placebo p of 0.51 and a like-for-like limit-order placebo p of 0.20 (one-sided, 200 replicates). A stricter variant that follows the walkthrough’s order (n 1,405) gives −0.021R gross [−0.117, +0.075] and −0.171R at 1 point.

Two descriptive checks. The first target was hit on 30.95% of trades (450 of 1,454), against 31.3% for a driftless random walk (30.2% when it starts at the CE): chance level. And ICT says his setups generally come with 3:1 and 5:1 built in by default (P5 18:07). In this recipe the first objective offered a median 2.25:1, and 32.5% of setups offered at least 3:1; the runner’s objective offered a median 6.6:1 but existed on only 30.6% of trades.

Not coded. The displacement fair value gap that, ICT says, takes greater command over price than the first one (P5 37:29); salience; inversion fair value gaps; and the full first-presented-fair-value-gap rule set. That classification, including the displacement rule, is what he says critics who ran the numbers lack (P5 32:28, 37:29). These results do not reach it.

“To the tick”: the −1.0 projection and the fib ladder

In Part 5 ICT draws a fib between the prior session’s final print and the 09:30 opening price. Drawn from the open to the final print, its −1.0 projection marks the high of the morning session (27:22–27:50). Drawn from the final print to the open on the same morning, it marked an 11:12 low, which he calls “the exact low to the tick” (28:28) and notes was later swept by a lower low; he places that lower low on a further multiple of the gap (30:12–30:23). His example day falls after our data ends on September 4, 2026, so we could not check it. Our test scores the morning’s high and low, which matches his claim about the high.

We asked how often the morning (09:30–11:59) high or low lands on the −1.0 level to the exact tick, compared with the neighbouring multiples 0.8, 0.9, 1.1 and 1.2, on days with gaps of at least 10 points (n 2,186; levels on the 0.25-point grid). It happened on 10 days: 0.5% of them, about 0.9 a year, against 7.4 expected if −1.0 were no different from its neighbours (p 0.38). Within 1 tick it happened on 23 days, 1.1% [0.7, 1.6], against 0.9% for the neighbouring multiples, a difference of +0.15 pp [−0.33, +0.62]. Any advantage is at most 0.9–1.6 extra days a year.

That explains why chart examples are easy to find. A 16-level ladder of gap multiples (0.5 to 4 in half steps on both sides, the kind of “standard deviations” ICT refers to in Part 5) lands within 2 ticks of a morning high or low on 142 days, about 13.3 a year. The same ladder shifted to other phases, so that it means nothing, does so on a median of 166 days. Any ladder of levels does this about once a month, so a journal fills up with “to the tick” screenshots whether or not a level means anything.

The fair reading is that −1.0 works about as often as its neighbours. For how ICT builds graded ranges and projections in the first place, see his “Chain of Custody” lecture, turned into code.

The calendar claims: NFP week and September

NFP week. ICT says the 08:30 jobs report, non-farm payrolls (NFP), creates volatility (P1 06:42), and that he teaches new students to avoid trading on the Thursday and Friday of NFP week (P1 06:52–06:58). In Part 3 he says a brand-new trader should be done by 11:00 ET on the Wednesday of that week (P3 28:11–28:33).

On NFP Friday the 09:30–10:59 range, scaled by its trailing 20-day median, is +0.169 log [+0.081, +0.257] wider than on other Fridays: about 18% wider [+8, +29]. The p is 0.0005, it is positive in 10 of 11 years, and it rests on 118 NFP Fridays (BLS release calendar) against 415 other Fridays. This is claim 9, the only claim that survives the ten-claim correction (claim p 0.001, Holm 0.010).

NFP Thursday is +0.022 log [−0.063, +0.108], or −6% to +11%, p 0.59: normal. The claim covers both days, and Friday carries it. Scored on Thursday alone, nothing in the family would survive.

That supports the premise of his advice, not the advice itself. No setup did detectably worse on either day after correction. Silver Bullet recipe A on NFP Thursday shows a difference of −0.254R against other Thursdays (gross, stop-first, touch fill; permutation p 0.052, Holm about 0.41–0.47), and it reverses in the slice from 2025 to September 4, 2026: a lead, not a finding. The CE’s lead over the mirror is no different on NFP days than on the same weekday in other weeks (difference-in-differences, gaps ≥ 2 pt: Thursday +1.4 pp [−11.9, +15.2], Friday −0.4 pp [−14.1, +14.0]).

September. As a rule of thumb, ICT says, Septembers are usually bearish months (P3 12:56–13:00), and he hedges it at once as just a general rule (P3 13:21–13:27).

Over 98 years (1928–2025), September has been the S&P 500’s weakest month: −1.93 pp against the average of the other months (p 0.00056), and down in 54.1% of years against 39.0%. On the Nasdaq-100, 1986–2025, Septembers fell in 45.0% [31, 60] of years (other months 38.4%). Its average is −2.03 pp against the other months, p 0.065, Holm 0.52. The median is −1.15 pp, and without 2001 and 2008 the mean is −1.30 pp.

We scored the claim on the Nasdaq-100 because NQ tracks it, and we made that choice after the results were known. It changes the verdict: scored on the S&P 500, the September claim would survive the correction (Holm 0.0056).

He also calls a November seasonal low a rare occurrence (P3 13:06). November actually holds 26.5% (S&P) and 20.0% (Nasdaq-100) of the lowest closes between September and November, which is no rarer than chance.

Very large gaps: the one exploratory leadexploratory

In Part 2, on a day with a gap of about 400 points, ICT says:

In the same breath he hedges: it may be a moot point, because the market could be very weak and simply keep selling off (26:42–26:48). He is describing a trade attempt, and adds that on days the attempt is stopped out, that is okay (26:57–27:02). What we can test is the premise behind it: whether price tends to make an attempt back into a very large gap.

One pattern resembles that remark. It shows up in-sample on NQ opening gaps of at least 1.355% of price, the size of his example day: 212 days since 2016. On those days price traded one-eighth of the way back toward the prior close by 10:00 on 78.3% of days, against 62.3% for the same distance away: +16.0 pp [+5.9, +25.7], a nominal figure found after the fact. We found it while searching 13 gap-size cuts × 8 levels. Across all of them, the gap’s direction does carry information about which nearby level trades first. But for ICT’s threshold the evidence is borderline once the search is counted: adjusted p 0.076 for touch, and between about 0.01 and 0.16 depending on the measure.

It is not the CE. On these days the CE and the mirror were each touched on 11.3% of days (24 against 24).

We then ran follow-up checks on other data (table below). On the Dow 30 CFD, our main follow-up market, it pointed the same way but was not significant (94 days). On the S&P 500 CFD, the secondary market, it was clear, but 91 of its 104 days were the same shock days as NQ’s, and 2020 and 2022 hold 70 of the 104.

On the only earlier NQ data (2010–2015), just 23 days qualified. Our verdict for that case: same sign on the earlier data but not significant (n ≈ 23, underpowered); the pattern remains an exploratory, post hoc lead. The trade’s minimum detectable effect on those 23 days was about 0.54R. Across stocks gapping at least 1.355% with NQ the pattern was small and not significant; only the 19 Nasdaq-100 heavyweights showed part of it, +6.4 pp [+3.5, +9.3] (post hoc).

G/8 rows: the level one-eighth of the gap toward the prior close, minus the same distance away, touched by 10:00. The Dow and S&P rows are Dukascopy bid CFDs. The NQ 2010–2015 row is the trade, net of 1 point, stop-first, on a 15:59 reference.

As a hypothetical trade, in-sample and post hoc. A mechanical version bought or sold at the 09:30 open toward the prior close, with a 1:1 bracket one-eighth of the gap away (so 1R is one-eighth of the gap; target fills on touch, stop at its level with no slippage), and went flat at 10:00. After a 1-point round-turn cost the backtest gave +0.220R per trade [+0.09, +0.35] (stop-first and neutral; +0.230 target-first), n 212. That figure is before any correction for the search.

It is fragile. 2020 and 2022 supplied 62% of the R from 42% of the trades; without them it is +0.147R (stop-first, t 1.66). With 2 points of entry slippage and 2 points of stop slippage it is +0.111R (stop-first, t 1.60). Opening fills on days with gaps of 1% or more are unmeasured: on March 16, 2020 price reopened 102.5 points beyond the stop, a loss of −2.60R rather than the modelled −1.02R.

We also checked how much of this in-sample result depends on the 09:30 print, by delaying the hypothetical entry one minute. The two ways of modelling that give different answers. With ICT’s 09:30-anchored levels kept and entry at the 09:31 open, the backtest gave +0.309R per G/8 (one-eighth of the gap) gross (+0.269R net of 1 point), against +0.353R per G/8 gross for the 09:30 fill; R is measured per G/8 here because the later entry changes the actual risk. That is on the 156 days not already decided inside the 09:30 minute (the other 55 are skipped), with the trade flat at 09:59; all tie rules give the same result. With the bracket re-centred on the 09:31 price instead, the net at 1 point falls from about +0.22R to +0.127 / +0.136 / +0.136R (stop / neutral / target, n 211).

These are hypothetical, in-sample backtest figures for a pattern found by searching many settings. They do not reflect real trading, and past or simulated performance does not indicate future results.

We treat it as a lead, not a strategy, and nothing here is a trading signal or a recommendation. About 20 days a year clear the threshold, so only several years of new data can settle it.

All ten claims, one correction

Testing ten claims at 5% each makes at least one false “hit” fairly likely. Holm keeps the chance of any false “significant” result across the ten at or below 5%. Each claim gets one statistic, chosen by stated rules (see Comparison 3 above).

Two-sided p-values. Claim 6’s “7.4 expected” is how many exact hits −1.0 would get if it were no different from its neighbouring levels. Claim 9’s p is twice the smaller of its two cells, Thursday and Friday (Friday p 0.0005).

One claim survives: claim 9, carried by NFP Friday. The other nine non-rejections are not evidence for ICT. Claim 7 has the opposite sign to his prediction, claim 3’s estimate is negative, and claims 8 and 10 point his way but do not survive.

Three choices change a verdict. Scored on the S&P 500, claim 8 passes (Holm 0.0056). Claim 10 uses the Dow, our main follow-up market; with the S&P instead it would pass, mostly on NQ’s own days. And scored on NFP Thursday alone, claim 9 gives nothing.

Outside the family, the literal 70% fails decisively: 57.7% against 70%, p 2.2 × 10⁻⁴⁰.

What a new student’s journal would actually show

ICT’s log trains the eye in hindsight. What the comparisons add is which recorded regularities can be told apart from what any level, any ladder or any random entry would produce.

ICT asked students to count. On ten years of NQ the count is 57.7%, and a level the same distance from the open in the opposite direction gets 56.1%. Nothing here says the CE is useless as a reference line in a journal, but on NQ since 2016 it does not behave like a magnet. Most of “about 70%” is how far NQ normally travels in 30 minutes on small gaps.

A 12-week log cannot separate 57.7% from 56.1%. It took 2,636 sessions to pin the difference down to between −1.0 and +4.3 percentage points.

What this does not show:

• That ICT’s method fails. Only mechanical translations were tested.
• That the CE carries no information anywhere. Pre-2016 NQ shows it ahead of the mirror, and US stocks show a small generic fade.
• That ICT promised a result. He hedged the number and never promised profitability, though he did say the count would reveal a recurring phenomenon, which is why checking it is fair.

What held up: NFP Friday mornings are wider, which is a statement about volatility and not a trade, and September has been the S&P 500’s weakest month. The large-gap “attempt” remains an open, exploratory lead.

The general lesson is about method. A claim that a level is reached “X% of the time” needs a comparison level the same distance away in the opposite direction. A journal that logged the mirror next to the CE would show the mirror getting hit about as often, though only years of entries could measure the small gap between them. Our placebo test makes the same point for trades, and the other “70%” in this series makes it for win rates.

What we could not test