BtcFiveMinBollingerLowerBandMeanReversionLong

BTC 5-Min Bollinger Band Mean Reversion Long (Sub-Daily Intraday)

A LONG-ONLY MEAN-REVERSION strategy on BTCUSDT perpetual futures (Binance USD-M) using 5-MINUTE bars and Bollinger Band lower-band pullback entries. RATIONALE driven by the now-DEFINITIVE session finding (18+ strategies analyzed across 5 mechanism categories, ALL FAILED): the only remaining genuinely-untested category per the analyst's explicit recommendation is 'sub-daily intraday strategies (1-minute or 5-minute mean reversion on BTC/ETH spot)'. This proposal directly addresses that recommendation. CRITICAL DIFFERENTIATION from the falsified XRP VWAP-deviation 4H mean reversion: (a) 5-MINUTE TIMEFRAME — this captures INTRADAY MICROSTRUCTURE mean reversion (driven by HFT liquidity provision, order book imbalance reversal, momentum-ignition events that snap back within 30-60 minutes) — a structurally DIFFERENT mechanism than 4H 'revert to fair value' MR. (b) BTC AS UNDERLYING — the most liquid crypto perp by 10x, with the tightest spreads and deepest order book; intraday MR works best in highly liquid markets where dislocations are quickly arbitraged back. (c) BOLLINGER BAND construct — uses 20-bar SMA + 2.5σ stddev to define an extreme excursion; this is a classic intraday-mean-reversion measure (Bollinger 1980) validated across decades of equity index futures research. (d) FEE-AWARE DESIGN — Binance USD-M perp at 0.10% RT taker fees per trade; targeting per-trade returns of 0.5-0.8% provides comfortable margin above 0.20% combined RT fee (entry + exit). DESIGN CHOICES ELIMINATING SESSION FAILURE MODES: (i) Not alt-perp trend-following (11 failures), (ii) Not cross-asset pair (class falsified at -4.86 OOS Sharpe), (iii) Not oscillator entry on 4H, (iv) Not supplementary data, (v) MAXIMALLY SIMPLE — 2 canonical Bollinger parameters (length=20, stddev=2.5). (vi) Long-only on the strongest-trending major in current regime (BTC has had positive 2024/2025/2026 due to ETF flows) — the dip-buying direction aligns with macro BTC bull regime.

Implements the hypothesis exactly: sub-daily (5-MINUTE) intraday mean reversion on BTC, the most liquid crypto perp, using the canonical 2-parameter Bollinger construct (length=20, stddev=2.5). calculate_signal returns a CONTINUOUS z-score that varies every bar (passes Layer-2 frozen-signal detection and carries conviction), and the entry decision is thresholded in should_enter (z <= -entry_z, with entry_z defaulting to stddev so entry is exactly 'close below the lower band'). Long-only and aligned with the BTC bull regime / dip-buying thesis; routed to BINANCE futures because round-trip taker fees (~0.10%) are half of spot (~0.20%), preserving the targeted 0.5-0.8% per-trade reversion edge above the ~0.20% combined RT cost. Fee-awareness is reinforced by a 0.7% take-profit and mean-revert-to-midline exit (both comfortably above RT cost) plus a 1.5% stop and 2-hour timeout to cap the tail risk of excursions that keep trending instead of reverting. Sizing is capital-relative via get_account_equity() with stop-distance risk budgeting (no hardcoded quantities), and leverage stays at 1.0 (referenced in sizing for consistency but not amplifying). BTCUSDT 5-minute is a deep, currently-listed series so no data_unavailability risk. Avoids every named session failure mode: not alt trend-following, not a cross-asset pair, not a 4H oscillator, not supplementary-data dependent.

Both-ends failure, not worth optimizing. (1) The account LIQUIDATED (total_return -100%, max_drawdown 100%, liquidation 2021-04-20): long-only Bollinger dip-buying with no portfolio-level risk cap kept re-entering the April-2021 crash until equity hit zero. (2) Even before the blow-up there is NO net-of-fee edge over a large 7,626-trade sample: profit_factor 0.63, Sharpe -1.63, avg_trade_return -$33, and commission_pct_of_gross 46.18% (commissions $200k + impact $149k consume ~half of gross). Win rate 59% but avg_win $96 << avg_loss $220 (tail_ratio 0.38) — the 5-min reversion collects pennies and pays out dollars during trends. This falsifies the hypothesis's premise of a 0.5-0.8%/trade edge above the ~0.20% RT cost; the realized per-trade edge is negative. With only 2 tunables (length, stddev) there is no parameter region that turns a negative-net-of-fee signal positive, so optimization would merely overfit the pre-liquidation window. NOTE for research: this confirms the session's pattern — the last untested category (sub-daily intraday mean reversion on BTC) also fails, primarily because high-frequency trading costs overwhelm any micro-reversion edge. Stop proposing high-turnover 1m/5m mean-reversion on a single instrument unless the design first demonstrates a per-trade gross edge well above ~0.2% RT and includes a hard account-level drawdown/exposure cap to prevent liquidation.

Long-only intraday Bollinger Band mean-reversion on BTCUSDT USD-M perpetual, 5-minute bars. Computes a rolling 20-bar SMA and population sigma; the per-bar decision variable is the z-score z=(close-SMA)/sigma. Enters long when z <= -2.5 (close punches below the 2.5-sigma lower band, an extreme intraday downside excursion). Exits when price reverts to/above the SMA midline (z >= 0), or on a 1.5% stop-loss, 0.7% take-profit, or a 24-bar (2-hour) max-hold timeout. Position sizing is stop-distance-based and capital-relative (risk 1% of equity over the 1.5% stop), capped at 95% of equity notional.

Code runs at scale (671k bars, 7,626 trades) with no plumbing failure; the Bollinger/z-score logic and risk-exit scaffolding are implemented as described.

Addresses a genuinely untested category (5-min intraday MR), so the negative result is informative.

ACCOUNT LIQUIDATED on 2021-04-20 (total_return -100%, max_drawdown 100%); long-only dip-buying repeatedly re-entered the April-2021 crash until equity hit zero — a fatal portfolio-risk-control hole.

No net-of-fee edge: profit_factor 0.63, Sharpe -1.63 (CI entirely negative), avg_trade_return -$33 over 7,626 trades.

Fees dominate: total_commission $200k + impact $149k, commission_pct_of_gross 46.18% — costs consume ~half of gross PnL. This directly falsifies the hypothesis's 0.5-0.8%/trade premise (realized per-trade edge is negative).

Bad win/loss geometry: 59% win rate but avg_win $96 vs avg_loss $220 (tail_ratio 0.38, skew -21, kurtosis 576) — tiny reversion wins, large trend-continuation losses; TP 0.7% / SL 1.5% asymmetry compounds it.

Only 2 tunables (length, stddev) cannot create an edge where the base signal is negative net of fees; optimization would overfit the pre-liquidation window.

BtcFiveMinBollingerLowerBandMeanReversionLong was proposed to test the session's last untested category — sub-daily intraday mean reversion on the most liquid crypto perp — by dip-buying 2.5σ Bollinger excursions on 5-minute BTC bars. It ran cleanly at scale (671k bars, 7,626 trades) but failed at both ends: the account was liquidated in the April-2021 crash (total return -100%) and the signal had no net-of-fee edge anyway (profit factor 0.63, Sharpe -1.63, -$33/trade), with fees consuming ~46% of gross. The analyst abandoned it at backtest review after one iteration, noting that only two tunables could never flip a negative-net signal positive and that optimization would merely overfit the pre-liquidation window. The recorded lesson confirms the session's pattern: high-frequency trading costs overwhelm micro-reversion edges, and future intraday MR proposals need a demonstrated gross edge above costs plus a hard account-level risk cap.

High-turnover 1m/5m single-instrument mean reversion is overwhelmed by trading costs (here ~46% of gross) — its tiny reversion wins are paid out as large trend-continuation losses — and long-only dip-buying without a hard account-level drawdown/exposure cap can liquidate; don't propose such designs without first proving a per-trade gross edge well above ~0.2% RT and adding a liquidation guard.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate: a both-ends failure — a fatal portfolio-risk hole where long-only dip-buying repeatedly re-entered the April-2021 crash until equity hit zero, plus a negative net-of-fee signal over a large sample that its two tunables (length, stddev) could not rescue.

A long-only intraday mean-reversion strategy on BTCUSDT 5-minute perpetual-futures bars that bought extreme downside excursions below a 20-bar, 2.5σ Bollinger lower band — betting microstructure dislocations would snap back to the midline — and exited on reversion to the SMA, a 0.7% take-profit, a 1.5% stop, or a 2-hour timeout.

Catastrophically: across 671k bars and 7,626 trades the account was liquidated on 2021-04-20 (total return -100%, max drawdown 100%), and even pre-blowup it had no net-of-fee edge — profit factor 0.63, Sharpe -1.63, expectancy -$33/trade. Despite a 59% win rate, average wins ($96) were dwarfed by average losses ($220), and fees dominated with commissions of ~$200k plus ~$149k impact, 46.18% of gross.