ποΈ Compression β Expansion (15-minute bars)
Backtest Β· 2026-04-20 β 2026-07-15 (60 trading days) Β· top-60 liquid names chosen from BEFORE the test window Β· built 2026-07-16 Β·
Trust Report
Verdict: NO EDGE β pass. The pattern fires constantly and loses a little bit everywhere: every tightness setting, longs and shorts, both halves of the period, and before costs.
The play tested: at least 5 fifteen-minute bars stacked inside a tight box (the coil) β the first bar that closes outside the box is the breakout β enter at that close, stop at the far side of the box. Long above, short below. Day trades only β everything closed by 4 PM. Two exit styles tried: take 2R, or ride to the close (results nearly identical).
| Coil tightness | Trades | Win% | Avg R | Median R |
| Tight coil (K=1.0) | 886 | 43.7% | β0.03 | β0.20 |
| Standard (K=1.25) | 2,031 | 44.6% | β0.04 | β0.16 |
| Loose (K=1.5) | 2,788 | 45.8% | β0.03 | β0.11 |
| Slice (standard setting) | Trades | Win% | Avg R |
| Longs only | 1,046 | 45.6% | β0.03 |
| Shorts only | 985 | 43.5% | β0.04 |
| First half of the period | 1,040 | 43.4% | β0.07 |
| Second half | 991 | 45.8% | β0.01 |
| Longer coils (7+ bars) | 100 | 45.0% | β0.07 |
| Entries 10:30β12:00 | 171 | 48.5% | +0.02 |
| Entries 12:00β1:30 (69% of all!) | 1,410 | 44.0% | β0.04 |
| Entries 1:30β2:30 | 450 | 44.7% | β0.04 |
π‘ Why it loses: 69% of these signals fire between 12:00 and 1:30 PM β coils mostly form at lunch, so this play is secretly a "trade the midday chop" machine, and midday chop breakouts suck traders in and die (44% win at 1:1 odds). The only slice that wasn't losing is 10:30β12:00 (+0.02R = breakeven, not edge). The coil isn't the fuel β it's just the quiet before more quiet.
See it on a chart:
β
TSM 2026-04-22 β worked (+2R)
β AMD 2026-04-20 β typical midday failure (β1R)
Green line = entry Β· red dashes = stop (far side of the box) Β· gold dashes = 2R target.
π§ Can it be fixed? 20 variants tested (2026-07-16)
Every classic "improvement" applied as a gate on the same 2,031 signals. Fishing-expedition rules: many comparisons β even a green cell here would need a forward test before anyone trades it.
| Improvement tried | Trades | Win% | Avg R | Verdict |
| Morning only (10:30β12:00) | 171 | 48.5% | +0.02 | breakeven β best cell, still not edge |
| Volume surge β₯1.5Γ on breakout bar | 261 | 45.2% | β0.04 | no help |
| Volume surge β₯2Γ | 90 | 37.8% | β0.17 | worse β the loud break is the trap |
| Big expansion bar (β₯1.2Γ ATR) | 160 | 42.5% | β0.06 | no help |
| Box sitting at the day's high/low | 329 | 39.8% | β0.15 | worse (t=β3.0) |
| Breaks to a NEW day high/low | 436 | 39.0% | β0.16 | worst, and statistically solid (t=β3.9) |
| Only with VWAP / with day's direction | 1,663 | 43.6% | β0.06 | no help |
| Longer coils (6+ bars) | 432 | 45.4% | β0.05 | no help |
| Wider target (3R), combos of the above | β | β | neg | no combo with nβ₯100 goes green |
πͺ The mirror test (the interesting part): since breakouts to fresh highs/lows lose β0.16R, we honestly simulated fading them (enter against the break, stop over the breakout bar, 2R target). Result: the fade loses too (β0.03 on new-extreme breaks, 32% win) β the initial momentum runs over the tight stop 2 times out of 3 before the reversal comes. Lunch-coil breaks chop up BOTH sides. The only fade cell at breakeven: day's-edge breaks (+0.00, n=320). Nothing tradeable either way.
The one lever this test could NOT check β and where the real hope lives: the universe. These were generic liquid mega-caps with no reason to move. James's playbook says expansion needs fuel: a catalyst, an earnings shock, SIP membership. The honest next test = same detector, but only on days the name was ALREADY in play that morning (provable-morning SIP list). If compressionβexpansion works anywhere, it's there β and if it fails there too, the pattern is fully closed.
β±οΈ "But people made millions on this pattern" β yes, on a different clock
We also built the play
the way the famous traders describe it (v2): only on names already moving Β±2%+ on heavy volume, coil holding the right side of VWAP, break in the trend direction, trailed instead of capped. Result:
β0.31R, n=71, both halves negative β even worse, because in-play names whip harder intraday.
The decisive number is
how far ANY coil break can run before the closing bell:
| Open profit reached intraday | How often (2,031 breaks) |
| +1R | 31.5% |
| +2R | 7.5% |
| +3R | 2.1% |
| +5R | 0.3% |
| Biggest single run in 60 days | +6.9R |
Darvas, O'Neil (VCP), Minervini, Zanger traded this pattern on
DAILY and WEEKLY charts, held for weeks to months β where a winner runs +10R to +50R. A breakout play that wins ~40% of the time NEEDS those rare huge winners to pay for the losers. On 15-minute bars the day ends before the tail can grow: the biggest winner available all spring was 7R, and only 1 break in 50 ever saw 3R.
The pattern isn't wrong β the clock is. Same coil on a daily chart = a swing trade (different lane, testable); on an intraday chart it's at best a timing tool inside an already-working play, never a standalone edge.
π
The DAILY-scale test (the documented version) β the pattern wakes up
Same detector on DAILY bars Β· 2024-07 β 2026-07 (2 years) Β· 300 liquid names ranked on mid-2024 data (pre-window, no survivorship pick) Β· longs only Β· exits: breakeven at +1R then ride until a close below the 20-DMA Β· ETFs purged.
| Cut | Signals | Win% | Avg R | Note |
| All daily coil breakouts | 1,041 | 31.6% | +0.10 | positive where intraday was β0.04 |
| Leaders only (30%+ prior move, near highs) | 53 | 24.5% | +0.27 | small sample |
| Leaders + volume dry-up in the coil | 28 | 25.0% | +0.71 | the full documented recipe β tiny sample |
| Non-leaders (control) | 988 | 32.1% | +0.09 | the gates do separate quality |
The tails finally exist β this is what was physically impossible intraday: best winners +13.7R, +11.9R, +11.5R, +11.0R; on leader signals 28% reached +2R open profit, 11% reached +5R, avg hold ~12 days. Compare intraday: biggest tail all spring was 6.9R and only 0.3% saw 5R.
β οΈ Read before falling in love: (1) win rate is 25-32% and the MEDIAN trade is a full β1R stop β you eat ~7 small losses to catch one runner, then must hold it for weeks; (2) the top-10 trades carry ~89% of ALL profit β miss a few monsters and there's nothing left; (3) year 1 made all the money (+0.19R avg) while year 2 was flat (+0.01R) β regime-dependent, so by our own both-halves rule this is PROMISING, NOT CONFIRMED; (4) the best cells (n=28) are too small for statistics. House rule applies: a gated 2-year backtest = hypothesis. It earns a forward watch lane, not live size.
π Honesty box: universe = top-60 by dollar volume ranked on 2026-03-02β04-17 data (before the test window) β no same-day cherry-picking. Fills modeled at 15-min bar closes (optimistic on fast breaks). No commissions/slippage β reality is slightly WORSE than shown. One trade per name per day, first signal only. Stops honored inside bars; gaps through the stop filled at the open (honest). Trade-by-trade log: 2,031 rows, kept on file.