BtcPerp15mLiquiditySweepReversalLS

BTC Perp 15-Minute Liquidity-Sweep Failed-Breakout Reversal (Stop-Hunt Fade, Long-Short)

A short-horizon, long-SHORT, single-instrument microstructure mean-reversion strategy on BTCUSDT.BINANCE perpetual futures using 15-MINUTE bars and OHLCV-ONLY data (deep multi-year history, no fragile supplementary streams). Core thesis is the liquidity-sweep / stop-hunt reversal: resting stop-loss and breakout orders cluster just beyond recent swing highs and lows. Aggressive flow (or a brief momentum impulse) frequently spikes price through an obvious swing level, triggers that clustered stop liquidity, and then — once the forced-order cascade exhausts and no genuine continuation follows — price snaps back inside the prior range. The strategy fades these FAILED breakouts: it does NOT chase the break, it waits for a same-bar rejection (price wicks beyond the level but closes back inside) and enters counter to the sweep, with a protective stop just beyond the sweep extreme. This is the inverse of the failed ADA candle-pattern strategy (which BOUGHT the local high with a tight in-trend stop and got chopped) and is deliberately run on BTC, the deepest-liquidity instrument where swing levels are the most widely watched and stop clusters are most real. At 15m on BTC, qualifying sweep-rejections occur multiple times per week (hundreds per year), giving >>10 entries per walk-forward window so the OOS windows and holdout are populated — directly addressing this session's recurring event-sparsity / all-zero-OOS abandons. Simple parameter set (4 core params) to resist the overfitting that killed the daily breakout and squeeze strategies.

Implements the stop-hunt-fade hypothesis exactly and inverts the failed in-trend breakout-buying mechanism: it does NOT chase the break — it requires a SAME-BAR rejection (wick beyond the level, close back inside) and enters counter to the sweep with a stop just past the sweep extreme. The 4 core params (swing_lookback, reject_ratio, stop_buffer, tp_r_mult) keep the surface small to resist overfitting. calculate_signal returns a CONTINUOUS, signed score every bar — intrabar close position mapped to [-1,+1] — which varies bar-to-bar (defeats the Layer-2 frozen-signal check) and whose sign matches the desired trade direction, while the strict sweep+rejection booleans (_long_setup/_short_setup, computed every bar, not gated behind a default-return) drive the actual entry in should_enter. Routed to BINANCE USD-M futures because the strategy is genuinely long+SHORT (CASH/spot cannot short) and because tight 15m sweep stops make leverage real: risk-based sizing on a ~0.1-0.5% stop yields >1x notional, so leverage=3.0 is consumed via the equity*leverage cap in position_size (self._leverage read from config.leverage) — not inert, so it won't be flagged leverage_set_but_unused. 15m BTC produces hundreds of qualifying sweeps/year, giving >>10 entries per walk-forward/holdout window to populate OOS and avoid the event-sparsity all-zero-OOS abandons. OHLCV-only and single-instrument — no supplementary or cross-asset dependency that could trigger data_unavailable.

Fundamentally broken — the strategy LIQUIDATED the account in ~3.5 months (liquidation_date 2020-04-16, total_return -100%, max_drawdown 100%, with impossible -109%/-203% single days from leveraged ~99.7% full-notional sizing). This is not a tuning problem: (1) the edge is strongly negative (profit_factor 0.42, expectancy -$32/trade across 5222 trades) with a doubly-adverse win/loss profile (41.6% win rate AND avg_loss 1.7x avg_win) — the opposite of the 2R design; (2) at 15m, commissions are 48.7% of gross PnL plus $69k impact, so costs alone bury the strategy; (3) the 'fade the failed breakout' thesis is refuted by the data — during the March/April 2020 trend, swept levels that closed back inside frequently CONTINUED, and fading them with leverage produced the liquidation. No parameter sweep can turn a PF-0.42, account-liquidating, fee-dominated strategy positive, and the optimizer cannot even run because the account is dead for ~95% of the window. FAILURE PATTERN: leveraged long/short intraday liquidity-sweep / failed-breakout fade on 15m BTC liquidates the account — the counter-trend fade carries unbounded risk when sweeps continue, the win/loss profile inverts to low-win/big-loss, and 15m commission drag (~49% of gross) erases any residual edge. Bidirectional sub-hourly mean-reversion fades on leverage are a recurring liquidation pattern this session (cf. the SOL 15m VWAP long/short variant) and should not be re-optimized.

Long-SHORT microstructure mean-reversion on BTCUSDT.BINANCE USD-M perpetual 15-MINUTE bars, OHLCV-only. Fades failed breakouts / liquidity sweeps: a LONG fires when a bar's low pierces below the prior 20-bar swing low but the close snaps back above it and finishes in the upper portion of the bar (close_pos >= reject_ratio); a SHORT mirrors on a swept swing high. The protective stop sits just beyond the sweep extreme (stop_buffer), take-profit is tp_r_mult (2R) of the initial risk, and a 16-bar time-stop closes stale trades. Position sizing is risk-based (0.75% equity risk per trade) divided by the tight sweep-stop distance, so notional naturally exceeds 1x equity and is capped at equity * leverage (3x) * 0.95. Long+short on futures.

Frequent signals as the hypothesis intended (9821 setups, 5222 entries) — event-sparsity is genuinely solved.

Clean microstructure logic (same-bar sweep + rejection, protective stop beyond the extreme) that runs without error.

ACCOUNT LIQUIDATED on 2020-04-16 (~3.5 months in): total_return -100%, max_drawdown 100%, with impossible single-day moves (-109%, -203%) signalling ruinous leverage at ~99.7% avg position.

Strongly negative edge: profit_factor 0.42, expectancy -$32/trade, win_rate 41.6%, avg_loss $95.7 vs avg_win $56.4 — low win rate AND bigger losers, the inverse of the intended 2R profile.

Fee-fragile: commission is 48.7% of gross PnL at 15m frequency, plus $69k market impact — costs alone overwhelm any edge.

Sizing misfires: 4441 setups dropped to zero size; the tight-sweep-stop risk sizing produces unstable, over-leveraged notional.

Post-liquidation the strategy is flat for ~95% of the window, so 3-phase optimization cannot run meaningfully.

BtcPerp15mLiquiditySweepReversalLS faded stop-hunt failed breakouts on deep-liquidity BTC 15m bars, explicitly engineered for frequent signals to populate walk-forward windows and avoid the session's all-zero-OOS abandons. It did generate thousands of trades, but the thesis was refuted: during the March/April 2020 trend swept levels kept continuing, and fading them with ~99.7% leveraged sizing produced a -100% wipeout in 3.5 months, atop a 0.42 profit factor, inverted low-win/big-loss profile, and 49% commission drag. The analyst ruled it fundamentally broken — another leveraged sub-hourly MR fade that liquidates, like the SOL 15m VWAP long/short — and abandoned it. It ended after one iteration as abandoned, never advancing to optimization or risk review.

Leveraged bidirectional sub-hourly mean-reversion fades are a recurring liquidation pattern: fading failed breakouts carries unbounded counter-trend risk when sweeps continue, the win/loss profile inverts to low-win/big-loss (the opposite of the intended 2R), and 15m commission drag (~49% of gross) erases any residual edge — solving event-sparsity by trading more does not help if each trade has negative expectancy and uncapped tail risk.

It was abandoned at the pre-optimization backtest-review gate (verdict: abandon) as fundamentally broken — a strongly negative, doubly-adverse (low-win/big-loss) edge with the fade thesis refuted (swept levels frequently continued during the 2020 trend), leveraged counter-trend risk that liquidated the account, and ~49% fee drag — so optimization (impossible with the account dead ~95% of the window) and all later stages were never reached.

A long/short single-instrument microstructure mean-reversion on BTCUSDT.BINANCE perps (15-minute bars, OHLCV-only) that faded liquidity-sweep failed breakouts — entering counter to a same-bar rejection where price wicked beyond a 20-bar swing high/low but closed back inside, with a protective stop just past the sweep extreme and a 2R take-profit — deliberately run on deep-liquidity BTC at high frequency to populate walk-forward windows and dodge the session's event-sparsity abandons.

It liquidated the account in ~3.5 months: total return -100%, max drawdown 100% (liquidation 2020-04-16, impossible -109%/-203% single days from leveraged ~99.7% full-notional sizing), with a strongly negative edge across 5,222 trades (profit factor 0.42, expectancy -$32/trade, 41.6% win rate, avg loss $95.7 = 1.7x avg win $56.4) and 15m costs that alone buried it (commissions 48.7% of gross plus $69k impact); sizing also misfired, dropping 4,441 setups to zero size.