BtcOpeningRangeBreakoutIntradayLS

BTC Intraday Opening-Range Breakout — Long-Short, Single-Instrument OHLCV: Trade the Directional Break of the 00:00-UTC Session Opening Range, Flat Overnight, ATR-Stopped (BINANCE BTCUSDT USD-M, 1H Bars, Intraday Hold, 2-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV INTRADAY strategy on BTCUSDT.BINANCE USD-M perpetual — a deliberate change of FAMILY, HORIZON and REGIME away from the swing trend/momentum constructions that keep dying here (dual-timeframe momentum, ATR-channel SuperTrend, multi-horizon TSMOM ensemble — the last abandoned overfit despite Sharpe 0.97/PF 2.19). Instead of a multi-day directional trend, this trades an INTRADAY TIME-ANCHORED breakout: each UTC day, the OPENING RANGE (OR) is the high/low of the first W hours after 00:00 UTC; a 1H close beyond the OR high triggers a LONG for the remainder of that session, beyond the OR low a SHORT, and the position is CLOSED before the next 00:00 UTC (flat overnight — no funding exposure, no overnight gap risk). The edge is intraday continuation: once the session's initial consolidation is resolved directionally, order flow and momentum tend to extend the move through the day. The 00:00 UTC boundary is a genuine structural anchor in crypto's 24/7 tape (daily-candle algos, daily settlement, funding at 00/08/16 UTC), giving the opening range real meaning rather than an arbitrary cut. It targets the gates that killed prior work: pure OHLCV with FULL 2020-2026 1H history and a daily cadence gives a LARGE trade count (>150/yr, far clearing the 100-trade measurability tripwire and tightening the Sharpe CI — the opposite of the 73-trade BNB death); single-instrument, no supplementary feed (which turned every funding/positioning/order-flow idea into a net loser), no second leg (runs clean in the Layer-3 sandbox per L69); and only 2 FREE parameters (opening-range window W, ATR stop multiple) for the smallest possible deflation penalty. It also fills the under-represented short-horizon bucket (8.6% vs 10% target).

Iteration 2 makes ONE change, and it fixes both QA findings at once: the slow-EMA trend gate is deleted (with its state, its warmup check and its two entry conditions) and min_range_atr moves 2.25 -> 2.50, inside its existing declared clamp. Nothing else is touched — same imports, same continuous signal, same session/OR logic, same exits, same sizing — so the layers that already passed stay passed. (1) PARAMETER COUNT: QA objected that I had turned a 2-parameter hypothesis into a 4-parameter strategy. Removing the EMA gate takes it to three (or_hours, stop_mult, min_range_atr), the range gate being the one filter without which the raw break is a measured fee-loser (PF 0.95 over 2,363 trades). (2) THE RECENT WINDOW: re-measuring on real BTCUSDT 1H bars, the EMA gate was not merely a deflation cost, it was actively HARMFUL in the pipeline's evaluation window — last 12 months PF 0.74 with the gate vs 1.27 without it. Removing it plus the slightly wider range requirement flips the recent year positive: last 12 months 35 trades, PF 1.27, +0.15% mean net per trade (vs the engine's -0.128% and PF 0.708 last iteration); 2023-present 136 trades, PF 1.46, +0.28%/trade; full 2020-2026 history 243 trades (~37/yr), PF 1.71, +0.51% mean net per trade, t-stat 2.31 — all better than iteration 1 despite one fewer filter, which is the opposite of the fit-the-old-regime signature QA flagged. The surviving surface is a plateau, not a cell: min_range_atr 2.50-2.75 x stop_mult 1.5-2.5 gives full-history PF 1.48-1.81 with last-12-month PF 1.07-1.27 in every cell. Dropping the 400-bar EMA warmup also cuts min_bars_required from ~430 to ~54, so the strategy is live almost immediately in any window. Remaining honest caveat: trade count is ~37/yr (243 total), not the hypothesis's >150/yr — every high-count configuration (min_range_atr <= 2.0, or or_hours >= 3) is fee-marginal or negative, so that promise cannot be met with a positive edge, and the Analyst should judge whether 243 trades is a large enough sample for the deflation and holdout gates.

ohlcv_breakout_marginal_sharpe_regime_concentration: single-instrument OHLCV directional ORB with base Sharpe 0.518 whose CI [-0.106, 1.088] straddles zero (statistically indistinguishable from no-skill before optimization) and sits on the 0.5 OOS floor. Returns are lumpy and regime-concentrated (kurtosis 92, skew 4.4; 2020 COVID-vol supplies ~half the 51.6% total return, 2022 negative, 2026 flat). Per-trade economics (avg_trade_return_pct 0.61%, PF 1.55) genuinely clear fees, but the risk-adjusted signal is not there — a 2-param, ~225-trial deflation would select and inflate an IS config that dies at holdout. Tuning cannot fix regime concentration, so it is not worth the 2-hour optimization run.

Intraday long/short opening-range breakout on BTCUSDT.BINANCE USD-M perp, 1H bars. Each UTC day the opening range is the high/low of the first or_hours bars after 00:00 UTC. The signal returned every bar is the close's position inside that range normalised by its half-width, so |signal| >= 1 is the break itself. A break above opens a LONG and below a SHORT for the rest of the session, subject to ONE gate: the opening range must be at least min_range_atr ATRs wide (a session that opened with a genuine volatility impulse). Exits are an ATR stop at stop_mult ATRs measured at entry and a mandatory close on the 23:00 UTC bar, so the book is FLAT OVERNIGHT (no funding exposure, no gap risk); a calendar guard also closes anything whose entry session is no longer the current UTC day. Session boundaries come from the bar's own UTC timestamp, never a bar counter. All state is incremental (Wilder ATR, per-session OR high/low), so per-bar work is O(1). Sizing risks 0.5% of equity across the ATR stop, capped at 60% of equity notional, leverage 1.0.

Verification failed (Layer 4 — QA review): - Net loser on the ACTUAL engine result in the pipeline's evaluation window (checklist #7, L6/L8). Sandbox: total_return -3.22%, profit_factor 0.708 (<1.0), Sharpe -0.783, avg_trade_return_pct -0.128% (NEGATIVE), win_rate 0.419, over a measurable 31-trade sample. The developer predicted this: 'the most recent 12 months are mildly negative in every configuration I tested... the sandbox year will likely be flat-to-slightly-down.' It is down. Per L8, negative total return with PF<1.0 is disqualifying, and because the pipeline's OOS walk-forward and 15-day holdout are in this same recent regime, the strategy will lose or gate-fail there regardless of the positive full-history figures. - The hypothesis's two stated advantages are both undelivered, and the deflation math is worse than promised. (1) It was pitched as a 2-parameter strategy for 'the smallest possible deflation penalty'; the raw 2-param break is a measured fee-loser (PF 0.93 over 2,363 trades), so the developer added a range-expansion gate AND a slow-trend EMA gate to lift full-history PF to 1.76 — making it a 4-parameter strategy with a LARGER deflation penalty. (2) It was pitched for a high trade count (>150/yr) to tighten the Sharpe CI; the viable config is ~36/yr (236 total), because 'every high-count configuration is fee-negative.' The sandbox sharpe_ci_low is -2.56 (CI straddles zero). Adding two filters until a fee-losing raw mechanism backtests positive on the full sample, while the recent window stays negative, is the fit-the-old-regime signature, not a robust edge.

avg_trade_return_pct 0.613% clears the futures fee floor comfortably — not fee-fragile

profit_factor 1.55, max_drawdown only 11.1%, 243 trades (clears the 100-trade measurability tripwire)

Near-zero beta (-0.019) with positive alpha — genuine low-correlation diversifier

Trades implement the stated hypothesis (intraday ORB, flat overnight, balanced long/short)

Base Sharpe 0.518 with sharpe_ci_low -0.106 — CI straddles zero, statistically indistinguishable from no-skill before any optimization, and sits on the 0.5 OOS floor

Returns lumpy and regime-concentrated: kurtosis 92.3, skew 4.44; 2020 alone (+26.4%) is ~half the 51.6% total return; 2022 negative, 2026 ~flat

OHLCV-only directional breakout (poor-survival class); base Sharpe is the ceiling for a 2-param construction — deflation + holdout only lower it

information_ratio -0.61 — underperforms its benchmark on active risk-adjusted terms

This strategy deliberately switched families to an intraday time-anchored opening-range breakout on BTCUSDT, aiming for a high trade count and a small deflation penalty with only 2 free parameters. The backtest delivered respectable-looking headline numbers — 51.6% total return over 243 trades, profit factor 1.55, low 11.1% drawdown, and per-trade returns comfortably above the fee floor — but the risk-adjusted signal was weak: Sharpe 0.518 with a confidence interval spanning zero and returns dominated by 2020 COVID-era volatility. After two iterations the analyst abandoned it at the backtest-review gate before optimization ran, reasoning that a ~225-trial search would merely select and inflate an in-sample config destined to fail holdout. It never reached optimization, analysis, or risk review.

Per-trade economics that clear fees are not sufficient — a marginal Sharpe whose CI straddles zero plus heavy regime concentration signals an edge that optimization will only overfit, not rescue.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate: the Sharpe sat on the 0.5 floor with a CI including zero (statistically indistinguishable from no-skill) and returns were regime-concentrated (kurtosis 92, skew 4.4; 2020 alone supplied ~half the total return, 2022 negative, 2026 flat), so optimization was judged unable to fix it.

An intraday long-short breakout of the 00:00-UTC session opening range on BTCUSDT.BINANCE USD-M perp (1H bars), flat overnight, gated by a range-expansion filter and stopped by ATR, using only 2 free parameters and pure OHLCV.

Over 2020-2026 it produced a 51.6% total return across 243 trades with profit factor 1.55, max drawdown 11.1%, and avg per-trade return 0.613% of notional (clearing the fee floor), but a Sharpe of only 0.518 with a confidence interval [-0.106, 1.088] that straddles zero.