---
name: reversal-zones
description: Reversal-zone study (2 Oct 2026): order blocks, golden pocket, fibs, range tops, EMAs, ATH drawdown, RSI, double bottoms tested vs matched random levels on 128 coins, 1h
sources: [chat]
aliases: [support study, where to buy, order block test, golden pocket test, reversal study, buy zones]
---
- [stated] 2 Oct 2026 Ben asked: what is real support (range top, order blocks 1h/4h/1D, golden pocket 0.618-0.66 of a move or whole coin, MAs, % pullback, 4h oversold), find high-probability reversal zones, check method, 3 iterations, 20 improvements.

## Method (final, iteration 3)
- Data: 1h candles, 128 coins with 300+ hours (RH 42, BSC 32, Base 32, Solana 21), GeckoTerminal, to 2 Oct 2026. Not 1000 coins: the other studies' data is not reachable from chat.
- Every zone built causally (only known at that hour). OB = last red candle before an impulse that breaks a confirmed swing high, move at least 2 ATR; +FVG variant. GP/fibs on swing low to confirmed swing high, moves 50-100% (mid) and 100%+ (big), 1h and 4h. Range = 48h/120h box under 50% wide, then close 10% above.
- Trade test: limit buy at zone top, stop -15%, target +30%, 7 days; conservative (bar touching both = loss, fill bar only counts the stop). Also 10/20 and 25/75.
- CONTROL that matters: same coin, same start time, a random level 3-15% above or below the real zone. Edge = zone win minus control win. Iteration 1 (coin placebo) and 2 (time-matched mid-bar entries) were biased: placebo itself scored +13 pts in iteration 2, so those numbers are void.
- Bootstrap CI by coin, coin-balanced, early half vs late half.

## Results vs matched control (15/30)
- Order blocks 1h/4h (+FVG): -3 to -5 pts, n=4,130. 1D OB +10 but n=36. KILLED.
- EMA 20/50 touches from above: -11 to -15 pts. EMA200 ~ -2. KILLED.
- Golden pocket 0.618-0.66: 0 to +8 pts (1h big 0.00, n=1,098). Not special. Whole-coin GP +4 (n=121).
- Fib 0.886 of a 100%+ move: +12, CI 0 to +23, but coin-balanced only +3. TESTING.
- Range top retest: 0 (n=58). Old high retest +6 (n=3,051). Double bottom +3.
- RSI 1h<30 vs random closes within 2 days: +9 (n=1,378, CI +6 to +11, both halves). Above EMA200: +12 (n=212).
- Depth from ATH: 50-70% down gave best payoff, +0.30R per trade (CI +0.17 to +0.44), 85%+ down was worst (-0.15R). Depth is real context; exact line is not.
- Best combo: old high retest with RSI 1h under 35: 45% win, +0.60R (CI +0.21 to +0.95), n=489, 61 coins.
- Confluence (2+ zone types at same price/time) did not beat single zones for OB/EMA; helped fib 0.786-0.886 (+0.36R vs -0.37R) and ATH-drawdown zones.

## v2, 2 Oct 2026: 177 coins (BSC 69, Base 44, RH 42, Sol 21), 2% round-trip cost added
- Old high retest = a coin breaks 15%+ above an earlier swing top, then falls back to that old top; buy there.
- Old high retest + RSI 1h<35: +0.33R after costs (CI +0.04 to +0.61), n=734, 83 coins, both halves +0.38/+0.29; works RH (+0.64R) and BSC (+0.42R), fails Solana (+0.01) and Base (-0.10); only +6 pts vs shifted-level control.
- Any level zone 50-70% off ATH: +0.20R (CI +0.04 to +0.37), both halves. 85%+ off ATH: -0.27R.
- After costs: order blocks -0.22R, EMA touches -0.23R (n=5,768 / 3,060), RSI 1h<30 alone ~0R.
- Files: rz_ev_0/1/2.pkl, rzshard.py, rzbest.py, rz_best.txt, fetchH3.py in the same outputs folder.

## v3, 2 Oct 2026: Ben's 61.8 plan vs old-high plan, 275 coins (98 brand new: TON, ETH, Sui, Arbitrum, more Base/Sol)
- Ben's rule: coin pumped 3x+, first touch of -61.8% from ATH, limit buy, target back to -38.2% from ATH, 14 days. Stop -75%: +10.2% per trade, win 42%, n=131. Stop -80%: +16.2%, win 53%, n=129 (better). Target +30%: +5.6%. Adding RSI<35: n=21, no help.
- Random entry with the same wide stop/target: +5.3% (survivor tailwind), so the real edge is about +5 to +11 pts.
- By chain: RH +13.9%, Sol +17.8%, BSC +16.7%, Base -4.9% (n 15 to 62).
- New 98 coins: Ben's ~0% (n=24, too few to judge); old-high + RSI<35 plan +6.4% per trade (CI +2.4 to +9.5, n=1,109) vs random -0.4%, so it held out of sample.
- Bug fixed on the way: a fill bar that gaps below the stop must count as flat, not a win (it inflated Ben's average to +126%).
- Files: cmp618.py, cmp618.txt, raw_1h_573pools.tgz, pools5.json.

## v4, 2 Oct 2026: Moneytaur-style order blocks (Ben's MOO screenshot)
- [stated] Ben's method (Moneytaur): untested 4h OBs, bodies not wicks, nested lower-timeframe OBs inside reinforce, odd timeframes like 7h, fresh OBs give strong bounces rather than reversals.
- Test: body OB = last down candle before a close above a confirmed swing high with a 2 ATR move; first touch only; 4h/7h/12h built from 1h; nested = untested 1h/2h body OB inside; outcome = 24h max bounce and +20% before -10% (72h); control = same activation, level shifted 5-20%.
- 275 coins: every variant bounced less than the control (4h 1,297 touches: +5.1% vs +7.4%; 7h 611: +5.3% vs +7.1%; nested 3+ 963: 14% vs 18% wins). Not tested yet: liquidity sweep, FVG, HTF trend filters from his exact checklist.
- MOO 0xD9dB30BB0D2b8d2eae3826A1372117E058791e18: detector finds the same untested 31 Aug 4h body OB at about $3.29M to $3.71M mcap that [[trading-edge-hub]] already lists.
- Files: obtest.py, obtest.txt, ob_rows.pkl.

## Caveats
- Universe = pools active now, so absolute R is survivor-biased upward. Edge vs control is not (same coin, same time).
- Files: /mnt/user-data/outputs/breakouts-2026-10-01/ (chat only): raw_1h_350pools.tgz, rz.py, rzrun.py, rzrep2.py, rzimp.py, rz_ev.pkl, rz_iter2.txt, rz_iter3.txt, rz_improve.txt.