ADA OKX 10-Day Channel Break with a 12-Hour Clock: Long-Short Event Trade, Flat ~91% of the Time (ADAUSDT.OKX 1H, pure OHLCV, 1x)

Single instrument, single leg, pure OHLCV, event-triggered trade on the OKX USDT perpetual swap ADAUSDT.OKX, using 1-HOUR bars (history on OKX runs from 2020-03). The trigger is a 1H CLOSE above the highest HIGH of the prior 240 1H bars (10 days), which enters LONG, or a close below the lowest LOW of the prior 240 bars, which enters SHORT. The position exits on a fixed 12-hour clock. The book is FLAT whenever no breakout clock is running: about 394 trades x 12h over ~55,000 hours, so ~9% time in market and ~91% flat. No regime gate, no supplementary feed, no ranking (L130). Pre-study (point-in-time, run by the research lead on the catalog's ADAUSDT.OKX 1H bars, 2020-05 to 2026-09, one event per 24h): n=394 events (~62/yr). Mean signed forward log return was +0.33% at 6h, +0.61% at 12h (t=2.5, hit 0.53), +0.57% at 24h, +0.70% at 36h and +0.78% at 48h. The per-hour curve reaches its first peak on a t-stat basis at 12h, which sets the clock (L166). Net of a 0.11% round trip, the 12h clock gives mean +0.50%/trade and PF 1.35. Years 2020/21/22/25/26 are positive and 2023/24 slightly negative (-0.16%, -0.18%/trade). Long side +0.73% (n=210), short side +0.24% (n=184). The same rule on BTCUSDT.OKX gave +0.22% at 12h and +0.43% at 24h (t=2.1), the same sign and weaker, which supports a mechanism rather than an ADA artefact. SELECTION DISCLOSURE (L122): this cell was chosen from a scan of 5 event mechanisms (6h/24h tail continuation, 72h breakout, 72h breakout after vol compression, 10-day breakout) x 2 symbols (ADA, BTC; the only liquid OKX 1H series with multi-year history in the catalog) x 5 clocks = 50 cells. The best cell's t of 2.5 is therefore NOT significant after that multiplicity, and deflation should treat it as a best-of-50. The venue choice (OKX) is driven by the corpus quota (OKX 0.2% of experiments) and by OKX having 6+ years of 1H ADA history. It is not a claimed venue edge.

Failure code: no_edge, taken from the first measured report per L159. The first report showed Sharpe 0.20, a bootstrap CI of -0.59 to 0.94 (straddles 0), PSR 0.60 and PF 1.21, with 2023, 2024 and 2026 negative. Its avg_trade_return_pct was 0.63% against a ~0.11% OKX round trip, so this is NOT fee_edge. The optimization then failed two HARD validity floors, neither waivable. (1) CPCV walk-forward is overfit: IS 0.99 vs OOS 0.39. (2) OOS Sharpe 0.386 is below the 0.5 floor. The holdout Sharpe is 0.061 on 251 trades (ratio 0.16). DSR is 0.729 against an expected-max Sharpe of 0.578 over 16 effective trials, and the result fails programme FDR (p=0.27). This on top of the hypothesis's own disclosed best-of-50 cell selection. The optimizer's chosen cell (entry_level 0.66, channel_bars 154) is no longer a channel breakout. It trades 3x more, pays modeled impact of 132% of gross, and its realized ledger is -15.9% with PF 0.96 and drawdown 34.5% against the pre-registered 25%. The pre-registered prediction met 1 of 5 checks. It is not beta: beta is 0.03. Under L126 re-optimizing the same mechanism is not an admissible iterate, and no parameter region in the sweep beat the 0.5 OOS floor. The 10-day channel-break 12h-drift on ADA is indistinguishable from multiple-testing noise. Separately, and unverified: the daily MTM series has recurring paired -X/+X prints that scale with equity (e.g. 2026-04-13/14) and an unreversed large negative print on the final bar (2026-08-14: -14.85% base, -84.7% optimized). These make metrics_reliable=false and should go to engine review. They do not change this verdict, which rests on the realized ledger and the CPCV/holdout results.

The signal is sound; the sizing is broken. The raw fill-to-fill signed move is +0.625%/trade (the pre-study predicted +0.61%), but modeled impact takes about 0.55%/trade because notional ($95k-$450k) exceeds capacity_usd of $66.6k. The book also keeps trading about $400k notional after equity hits 0, with the per-trade leverage field at 10-15x on a 1.0x config. Fixes: (1) notional = min(equity*notional_fraction, max_notional_usd), with a new parameter max_notional_usd defaulting to 25000. (2) Add a hard guard in position_size: return 0 if equity <= 0 or notional/equity > config.leverage, and log get_account_equity() at each entry to find the stale equity read. (3) Leave channel_bars, hold_hours, cooldown and entry_level unchanged, and do not retune exits. The ONE metric expected to move is the ledger mean price_pnl_pct, from 0.07% to at least 0.40%. Impact should fall below 25% of gross, and total return should turn positive. If the ledger mean stays below 0.15% after the notional cap, abandon. Details are in workspace/discussions/f136b90e-8828-44c1-84c1-ce5711eafcae/iteration_1_feedback.md. QA's impact concern is borne out and is the whole story. QA's funding concern is moot, because funding is folded in this run. ## Library refinements (from the knowledge library; test them, do not assume them) The library backs the analyst's diagnosis: it caps order size in USD to limit slippage, and its turtle breakout code sizes each trade from a fixed risk budget divided by 2xATR rather than from all of equity. Two changes are grounded here: a hard notional cap with an equity guard, and ATR risk-budget sizing. A third, a 2xATR protective stop, is optional and should wait until the sizing fix has been tested. 1. [cost] Hard per-trade USD notional cap plus equity/leverage guard: In position_size: notional = min(equity * notional_fraction * config.leverage, max_notional_usd), with max_notional_usd = 25000 (well below the $66.6k capacity_usd). Return 0 if equity <= 0, or if notional / equity > config.leverage (1.0). qty = notional / close, rounded down to size_precision. Log equity and notional at every entry. Leave channel_bars=240, hold_hours=12, cooldown_hours=24 and entry_level=1.0 unchanged. — The signed move per trade was +0.625%, which matches the pre-study, but about 0.55% of it went to modeled impact because notional grew to $95k-$450k. The book also kept trading after equity reached 0. The arbitrage course's order parameters include a fixed per-order USD ceiling, trade_max_usd_every_time: '每次最多下单多少个USDT/BUSD的订单…防止对冲滑点过大' (cap each order's USDT size to stop slippage getting too large). The vn.py CTA docs likewise treat a separate target-position algorithm as the way to '拆分巨型委托，降低冲击成本' (split huge orders to cut impact cost). A fixed USD cap keeps impact roughly constant per trade while equity compounds. (source: 套利软件策略参数说明和使用 p.1; CTA策略模块 p.1) 2. [sizing] Turtle-style ATR risk-budget sizing instead of 95% of equity: Compute ATR(14) on 1H bars from bars that have already closed. Size each trade as qty = (risk_pct * equity) / (2 * ATR14), with risk_pct = 0.005 (0.5% of equity at risk per 2-ATR move). Then clip notional to min(equity * config.leverage, max_notional_usd = 25000). New parameters: atr_window = 14, risk_pct = 0.005, atr_stop_mult = 2.0. — The library's turtle breakout on hourly bars sizes each trade as trade_volume = risk_loss_money / (atr_value * 2), with the comment '1% - 2%' of capital at risk. ADA's hourly volatility varies several-fold between 2020-21 and 2023-24. Sizing on volatility means the high-vol breakout years, which drive both the edge and the impact, do not take outsized notional. It also means a stale or inflated equity read cannot turn into 10-15x effective leverage, because the notional clip still binds. (source: class_10_turtle_strategy.py p.1; class_12_verify_and_optimize_strategy.py p.1) 3. [stop] 2xATR protective stop inside the 12h clock (defer until after the sizing fix): At entry, record stop = entry_close - 2.0 * ATR14 for longs, or entry_close + 2.0 * ATR14 for shorts. In should_exit, exit on the first 1H close beyond the stop. Otherwise keep the 12h clock exit. Only add this after the notional-cap run confirms the ledger mean is at least 0.40%. The analyst asked that exits not be changed in this iteration. — In the library's turtle code, every breakout entry places a fixed initial stop at once: long_stop_loss_price = long_entry_price - 2 * atr_value (and the mirror for shorts). This is separate from the exit channel. The current strategy has no stop at all, and at 0.95 notional_fraction a single adverse 12h window on ADA can take a large bite of equity. A close-based 2xATR stop keeps the clock-exit hypothesis intact and bounds tail losses. It will cut a few winners that recover within 12h, so our backtest must confirm it does not reduce the +0.6% mean. (source: class_10_turtle_strategy.py p.1; class_09_turtle_strategy.py p.1)