The Trial Log

The claim

Price probes above the prior day's value area high (or below the low), the auction fails, and the failure is faded back toward the prior day's point of control. The version actually traded: a limit order at the boundary on a failed probe, stop just past the probe extreme (capped at 3 to 4 points), first target +2 points, runner to the prior POC, stop moved to breakeven after the first target. Two contracts, entries 09:45 to 15:30 ET, two losses or $1,000 ends the day.

Result of the rule as written

• Window A: the filters fired so rarely that only 5 trades qualified in 66 sessions. Gross −$212. The POC target was reached 0 times.
• Window B: 36 trades, 47.2% winners, gross −$1,988. POC target reached 5 of 36 times.

Two years, two volatility regimes, both negative.

Why it loses

Two structural faults, neither of which is about picking the right boundary.

• The breakeven stop sits inside the noise. Entry ±0.25 is narrower than a single 1-minute ES bar (median RTH bar range 2.50 points, mean 3.14), so nearly every runner was stopped at entry before it could travel.
• The target is a day's work against an hour's risk. Average distance to the prior POC was 19 points against an average stop of 3.2, a 6.6R geometry. 79.5% of trades finished between −1R and 0.

The concept on its own

With every filter removed, each boundary probe was compared with a random entry in the same session, on the same side, over the same horizon.

• Window B: probes beat random by +1.98 points at 60 minutes (t = 6.23, n = 7,160) and +4.14 points at 120 minutes (t = 9.12, n = 5,714).
• Symmetric brackets of ±3, ±4 and ±6 points: probes won first-touch more often than random by 4.7, 2.8 and 3.3 percentage points (z = 5.87, 3.51, 4.04).
• Window A showed the same direction on a smaller sample (+2.50 points at 60 minutes, t = 1.68).

A survival table of fixed targets from 4 to 12 points on Window A crossed zero between 6 and 7 points and stayed positive above, best +0.28 points per trade at 12 (t = 0.76). The gross edge at an 8-point target, +0.52, clears friction of 0.34, but the distribution is heavily right-tailed (median favourable excursion 3.25 points, mean 9.05), so separating that from zero would need several thousand trades.

What would change the verdict

The probe-versus-random gap disappearing on a third, untouched year. Or a fixed-target version reaching significance on 2,000 or more trades, which would upgrade the rule from failed to survived.

Next

The 4-to-12-point survival table, run unchanged on Window B.

The claim

Read the first hour. Decide whether today is a trend day or a balance day. Trade the afternoon accordingly: go with the move on a trend day, fade the edges on a balance day. Every discretionary ES method has some version of this step, and so did mine.

Method

The morning (09:30 to the cutoff) was measured three ways: Kaufman efficiency ratio (net move divided by path length), a variance ratio, and the false-break rate (breaks of the prior 20-bar extreme that closed back through the level within 10 bars). The same three measures were computed for the trading window, and each morning measure was tested as a predictor of its afternoon counterpart. Morning direction was tested against afternoon direction. As a control, morning range was tested against afternoon range.

Result

• Efficiency, morning to afternoon: correlation −0.019 at 10:00, +0.005 at 10:30, −0.078 at 11:00. All p > 0.22.
• Variance ratio: +0.059 / +0.080 / +0.064. Null, and the second half of the year collapsed each time.
• False-break rate: −0.048 / +0.030 / +0.036. Null.
• Terciles at 10:30, chop / middle / trend mornings: afternoon efficiency 0.073 / 0.071 / 0.076; afternoon false-break rate 78.2% / 79.3% / 80.4%.
• Direction: morning direction matched afternoon direction 54.7% (10:00), 54.3% (10:30), 50.6% (11:00); unstable between halves. Signed morning move vs afternoon net move, rho −0.027, p = 0.67.

The control

Morning range predicts afternoon range at +0.568 (10:00), +0.501 (10:30), +0.541 (11:00), p < 0.0001, stable in both halves of the year. The pipeline finds a real relationship when one exists. Size carries over from morning to afternoon; character does not.

One weak signal, reported for completeness

Overnight efficiency against afternoon efficiency: rho +0.156, p = 0.015, but it decayed from +0.233 in the first half to +0.082 in the second. Prior-day efficiency: null (−0.019). Not treated as a finding.

What the instrument is

Between 10:00 and 14:00, on essentially every session regardless of morning character, ES has a mean efficiency of about 0.07 and a false-break rate of about 0.79. Roughly four of five breakouts fail. That is structural, not conditional. “Trend day” is a label that can be applied at 16:00.

What would change the verdict

Any morning feature predicting afternoon efficiency out of sample at p < 0.05 with both halves of the year agreeing.

Next

Fair value gaps, on tick data, with the definition fixed before the file is opened.

The claim

Price comes back to fill a fair value gap and reacts from it. Buy the retrace into a bullish gap, sell the retrace into a bearish one, stop past the gap, target 2R.

Method

Definition fixed before the data was opened. A bullish gap is three consecutive 5-minute RTH candles where candle 3's low is above candle 1's high (bearish: candle 3's high below candle 1's low), at least 1 point wide. Entry is a limit at the near edge on the first return, no later than 15:30; stop one tick past the far edge; target 2R; flat at the close. Fills simulated on 1-minute bars, stop assumed first when stop and target share a bar, one tick of slippage on stops and $5 round-trip commission. Each retrace into a gap was also compared with 100 random entries in the same session, same direction, same holding time.

Result

• Fill rate: about 75% of gaps were revisited the same session (77%, 74%, 75%).
• As traded, net per trade: −0.55 points (main window, 95% CI −1.32 to +0.20), −0.61 (−1.29 to +0.07), −1.10 (−2.04 to −0.12). Win rate 36%, 34%, 28%, against about 33% needed for a 2R trade before costs.
• Gap retrace minus random entry, at 15 / 30 / 60 minutes: −0.11 / −0.32 / −0.06 points (main), −0.72 / −1.29 / −1.41, +0.34 / +0.87 / −0.89. Seven of nine below zero, largest |t| 1.59.
• Sensitivity on the main window: minimum gap 0.5 or 2 points gives the same picture. On 1-minute bars (811 trades) the gap entries did worse than random, t −2.03 / −2.47 / −2.79 at 15 / 30 / 60 minutes.

Limits

ES only, RTH only, no higher-timeframe bias filter. This tests the building block, not the full model it is usually traded inside.

What would change the verdict

A written filter that turns the gap-versus-random number positive, at p < 0.05, in a window it was not built on.

Next

Killzones, with the time windows fixed before the file is opened.

Multiple evaluations over the year, a pass rate in line with the firm's published base rate, no better. The early attempts, traded on ICT concepts, passed none. A later volume profile method passed several and took payouts. Multiple funded accounts were still lost afterwards. The method changed completely between the two eras; the funded-stage failure did not. That is why the trials on this site test rules, and why the live account below tests the trader.