Aurelia @fullypriced

Update on this post (23 Oct). The code and the registration docs aren’t public anymore. The write-up says so now, with a dated note. None of the numbers changed. I also shouldn’t have written “small samples are small.” The wider 2022-model test used different rules (any of four levels, an earlier start), and the overnight levels in both versions didn’t match what I registered. So those two numbers aren’t really comparable. The result is the same: no cell met the rule I set in advance. fullypriced.com/claims/ict-sil…

I froze ICT's Silver Bullet rules before looking, ran them once on almost 7 years of MNQ, and the Silver Bullet hour didn't stand out from the other 19. No cell met the registered finding rule. Numbers and two disclosed departures below 👇 I froze the rules for ICT's Silver Bullet and 2022 model, then ran them once on MNQ 1-minute bars (2020 to Sep 2026) against a resampled random-path baseline. 44 cells, in-sample and out-of-sample. Silver Bullet, 10:00–11:00 NY, win rate vs baseline: In-sample: 33.8% vs 34.9% (z −0.68, 828 trades) Out-of-sample: 38.4% vs 35.7% (z +1.39, 560 trades) A finding needed z > 3 in-sample, z > 2 out-of-sample, a win-rate gap of at least 5 points in both, and positive net expectancy in both. The same rule in the other 19 hours, pooled in-sample: 35.25% vs 35.23%. The 10–11 hour ranks 19th of 22 in-sample, 5th out-of-sample. The 2022 model: 117 fills. Out-of-sample +$22.22 per fill, on 56 fills. Very screenshot-friendly. Small samples are small. A wider re-registered version: +$4.65 on 196 fills out-of-sample, +$2.44 on 492 over the whole period (z +0.21), net of costs. Descriptive only. Two departures from my own registration: 1. The baseline's calibration check failed as written (4 of 34 qualifying cells). I scored anyway, on a judgment. 2. 2022-model overnight levels ran 00:00–09:29, not the registered 18:00–09:29. Not tested: discretionary ICT, other instruments, other periods. The spec was frozen privately, not published first. Write-up, redacted registrations, code, every cell: fullypriced.com/claims/ict-sil…

@MrMilkTrading This year’s results could potentially work well for a prop firm account.

10/ Day leg: flat where you picked it, spectacular where you didn't. That's an out-of-sample discovery, not a confirmation — and reporting it spent the holdout.Also: "overnight didn't run >0.15 ATR" is the subset where the night leg didn't pay. The curves can't be added.

1/Unusually good thread — publishing the DSR, the PBO, the 616 trial count and the day leg's in-sample flatness is rare, and it's the part most people skip .You asked for feedback. One control problem, three missing tests. All runnable on data you already have.

The Milk Man @MrMilkTrading

a month ago

@LindaRaschke posted one line on ES this morning. A Taylor buy day setup. Then @wojtek_speaks pointed it out to me. I ran a quick test on 18 years of ES data -- this one needs more validation but as usual, Linda is on point! 🧵

4 5 77 10K 130

9/ On NQ — the t-stats are the vol-normalised comparison, and they're 2.00 vs 1.85. Same result. The pts/trade gap is NQ's volatility, not a better edge.And day-NQ at t=0.73, 43% win, n=14 shouldn't sit on a card next to a net profit figure.

8/ Which points at a fix rather than a kill: set the forward expectation at the median of the plateau and trade the ensemble, not the argmax.Also worth running: a block bootstrap by decline episode. 3-lower-close signals overlap, so effective n is well under 94.

7/ Missing test 3, and the one I'd most want: the median OOS across all 616. Not the neighbours — the median.PBO 0.28 + DSR 0.02 says the family may be real but the config has no claim over its neighbours. If the median pays too, selection didn't do the work.

6/ Missing test 2: drop-best-k.Same study, 748 nights — the best 5% of nights were 133% of net. Drop the best 10 and net halves; drop 20 and t falls to 0.39.Scaled to n=94 that's "drop the best 1.3". You have the trade list. One line, and it settles the question.

5/ Missing test 1: event concentration. An independent MNQ study of the same family — FOMC-eve nights were 3.6% of nights and 19.1% of net. Unconditional version: 3.3% / 25.5%. NFP-eve: nothing. At n=94, four such nights at ~6x weight is ~20% of your OOS result.

4/ Fix is in your own data: normalise by a causal 20-session median range, or just restrict the control to the years the signal actually fires in.Same issue in $10K vs $41K. As a share of index level it's ~1.1%/yr vs ~2.3%/yr — about 2x, not 5.6x.

3/ Your own numbers show it: $43.3K / 89 = 9.7 ES pts per trade in 2020-26, vs the 4.4 full-history conditional. An independent MNQ estimate puts the modern-era unconditional 18:00→09:30 nearer +1.5–1.9 ES pts than +0.4.Era-matched that's ~5-6x, not 11x. Real, half the size.

2/ The control is era-mismatched."+0.4 pts on every session since 2008" is an 18yr points average. But 3-lower-close streaks cluster in high vol, and high vol clusters in the back half of a window where the index rose ~6x.The signal sample isn't drawn from the same years.