Strategies

BTC Dual-Timeframe Momentum Confluence (Single-Instrument USD-M, Long-Short, 4H Primary + 1D Confirm, Flat on Disagreement, Trailing-Stop Winners)

Outcome: Abandoned

BtcDualTimeframeMomentumConfluenceLS_HL

Outcome Summary

BtcDualTimeframeMomentumConfluenceLS_HL tried to reproduce the factory's one promoted single-asset survivor by keeping its two-signal confluence with flat-on-disagreement structure while replacing the funding leg with a daily momentum leg, trading BTCUSD on Hyperliquid with an ATR trailing stop and only three tunables. The code ran cleanly and generated a decisive sample — 173 trades over 803 days with no dropped orders — but the result was negative in every calendar year: -5.17% total return, Sharpe -0.047, profit factor 0.923, and per-trade return of -0.109% versus the +0.15% floor, with the 93 shorts alone losing $8,293 net and the regime split inverting the trend premise. The analyst abandoned it at backtest review rather than spending optimization hours, noting that gross PnL before any costs was only +$1,136 and that two lookbacks and a trail multiple cannot manufacture information absent from 173 observations. It was the third Hyperliquid run of this structure to die with the same signature, and the recorded takeaway for research is that the collinear slow-momentum leg degenerates the gate into near-always-in TSMOM.

Strategy report

Hypothesis

A long-SHORT, SINGLE-INSTRUMENT, pure-OHLCV momentum strategy on BTCUSD.HYPERLIQUID USD-M perpetual futures that takes a directional position ONLY when a FAST (4-HOUR) momentum signal and a SLOW (DAILY) momentum signal AGREE, and stays FLAT when they disagree. It generalizes the WINNING STRUCTURE of the factory's one promoted single-asset survivor -- the ETH funding-confirmed trend (price-trend AND funding AGREE, flat on disagreement) -- by replacing its non-transferable funding leg with a PURE-OHLCV slower-timeframe momentum leg. This fuses the two facts this session validated: (A) DIRECTION = momentum/continuation. Crypto CONTINUES, it does not revert, at every horizon tested this session -- every fade died anti-predictively (VWAP-excursion-MR BTC -90%/SOL -80%, volume-flush-bounce -45%, OI-divergence-fade, cross-sectional 2-day REVERSAL -94%), while the correct-signed cross-sectional MOMENTUM basket is POSITIVE (Sharpe ~1) and in optimization. (B) STRUCTURE = two-signal confluence with flat-on-disagreement, the robustness mechanism that carried the ONLY promoted single-asset strategy. EVERY design choice respects this session's kills: (1) PURE OHLCV, SINGLE INSTRUMENT -- immune to the EIGHT supplementary-data coverage kills (OI/taker/L-S/liquidations all data-poor) and the basket data-collection wall (thin alts abort runs). (2) MOMENTUM, not fade -- the repeatedly-refuted anti-predictive fade family is closed. (3) CONTINUOUSLY-EVALUATED confluence, NOT a rare conjunction -- the two momentum signals agree often (in trends) and disagree in chop, so the gate fires regularly and never starves walk-forward OOS windows (the trade-starvation / zero-OOS-window death that killed the Bollinger-squeeze conjunction). (4) TRAILING-STOP payoff -- the flush-bounce died on a losing-by-construction 1:3.3 reward:risk; here a trailing stop lets confluent trends run for avg_win >= avg_loss. (5) NOT a funding-confirmed-trend clone (uses NO funding, so not subject to its ETH/BTC-only limit -- though BTC is chosen as a robust major), NOT single-TF TSM (the confluence + flat-on-disagreement twist), NOT VWAP/impulse intraday continuation (different mechanism/horizon than my pending siblings), NOT cross-asset ratio, NOT supplementary-primary. Intentionally minimal: 3 tunables (fast_lookback, slow_lookback, trail_atr_mult).

negative_expectancy on a reliable 173-trade / 803-day sample, and specifically a GROSS signal failure rather than a cost or sizing one. Verified on the detail file: price PnL before commission is only +$1,136 across 173 trades (~+0.06% of notional per trade) against $3,741 commission and $2,571 of credited funding paid — removing 100% of trading costs still leaves the mechanism unable to clear a 0.09% round trip plus an hourly funding clock on 3.6-day holds. The failure is directional hit-rate and is one-sided (LONG +0.408% vs SHORT -0.778% gross/trade; the 93 shorts lose -$8,293 net), regime attribution inverts the trend premise, and 5 trades supply +$18,378 against a book total of -$5,176. Not iterate: three tunables (two lookbacks, a trail multiple) cannot manufacture information absent over 173 observations. Not revise_hypothesis: this IS already the venue reframe and the third HL run of this structure to die identically. FOR THE RESEARCH LEAD: the structural argument — swapping the promoted ETH funding-confirmed trend's funding leg for a slower-timeframe momentum leg — is what keeps failing. Funding was an independent information source; a slower lookback of the same price series is nearly collinear with the fast leg, so the confluence gate adds lag rather than confirmation and the book degenerates into near-always-in TSMOM (79% exposure).

Implementation

Long-short single-instrument BTCUSDT.BINANCE dual-timeframe momentum confluence, 4H primary + daily confirmation, pure OHLCV. It goes long only when both the fast 4H momentum (sign of the 12-bar return) and the slow daily momentum (sign of the 20-bar return) are up, short only when both are down, and stays flat on disagreement. It lets confluent trends run via an ATR(14) trailing stop (3x ATR from the position's running extreme) and exits early only on a higher-timeframe (daily) regime flip against the position. Risk-based sizing (1% equity risk over the ATR trailing distance) capped at 50% notional, leverage 1.0. Three tunables: fast_lookback, slow_lookback, trail_atr_mult. Ported from BTCUSDT.BINANCE to BTCUSD.HYPERLIQUID (2026-09-17); same code, evidence re-earned on this venue.

Generalizes the structure of the factory's one promoted single-asset survivor (two signals AGREE, flat on disagreement) by replacing its non-transferable funding leg with a pure-OHLCV slower-timeframe momentum leg, fusing the two session-validated facts: direction = momentum/continuation (every fade died anti-predictively; the correctly-signed momentum is positive) and structure = two-signal confluence with flat-on-disagreement. The daily confirmation is supplied as a slower-timeframe extra leg on the SAME instrument via extra_bar_types; because its timeframe differs from the primary 4H it is not a sync leg, so the cross-leg alignment barrier never blocks the 4H signal, and on_extra_bar maintains the daily momentum sign. extra_instruments lists BTCUSDT.BINANCE index-paired with the daily extra_bar_type so the on-demand collector fetches it (no data_unavailable). calculate_signal returns the continuous fast momentum return so Layer-2 frozen-signal detection passes, while the discrete confluence decision (both timeframes agree) lives in should_enter. Exits are designed to let winners run for avg_win>=avg_loss (the fix for the flush-bounce's losing 1:3.3 geometry): an ATR trailing stop from the position's running extreme, plus an early exit only when the higher-timeframe daily regime flips against the held side (fast-TF chop tolerated). Sizing is risk-based (1% equity over the 3x-ATR trailing distance) capped at 50% notional, clearing the BTCUSDT $100 min-notional. The confluence gate fires continuously (agreeing in trends, disagreeing in chop), so walk-forward OOS windows are densely populated, avoiding trade-starvation. Single-instrument pure-OHLCV avoids the supplementary-data and basket data-collection walls; BTC is a robust major. Leverage stays 1.0, referenced only as a non-amplifying sizing multiplier, so the unused-leverage gate does not apply; min_bars_required guarantees the fast lookback and ATR are available before any trade, and slow_dir stays 0 (no entry) until the daily history is sufficient.

Backtest Review

Sharpe: -0.05
Total return: -5.17%
Max drawdown: 12.27%
Trades: 173
Win rate: 35.3%
Profit factor: 0.92

Code runs cleanly and implements the stated mechanism: 173 entries over 803 days, 80 long / 93 short, 3d 16h average hold, 79% exposure, no dropped orders.

Sample is adequate to measure: 173 trades, 754 daily observations, 2.2 years — a conclusive negative, not an unmeasured one.

Costs fully accounted: commission folded into price_pnl, funding credited (-$2,571 paid), impact $1,109 folded.

Negative expectancy: avg_trade_return_pct -0.109% vs the +0.15% HYPERLIQUID floor, PF 0.923, Sharpe -0.047, total return -5.17%.

GROSS signal failure: summed price PnL before commission is only +$1,136 over 173 trades (~+0.06% of notional each) vs $3,741 commission and $2,571 funding — zero-cost limit still fails.

One-sided: LONG +0.408% gross/trade vs SHORT -0.778% gross/trade; the 93 shorts alone lose -$8,293 net.

Top 5 trades = +$18,378 against a book total of -$5,176; 11 consecutive losses; 35.3% win rate.

Negative in all three calendar years; regime split inverts the trend premise (stressed tercile -11.25%, Sharpe -0.56); 764 of 803 days in drawdown.

Third run of this exact mechanism on this exact venue with the same signature (priors: -0.146%/PF 0.889 and PF 0.90/Sharpe -0.137).

Outcome Summary

A slower lookback on the same price series is nearly collinear with the fast leg, so it adds lag rather than independent confirmation — the promoted structure's power came from funding being a separate information source, not from two-signal confluence per se.

The analyst returned an abandon verdict at the pre-optimization backtest review for negative expectancy on a reliable sample — and specifically a gross signal failure, since removing all trading costs still left the edge unable to clear a round trip. Optimization, the post-optimization analyst stage and risk review were never reached.

A long-short, pure-OHLCV BTCUSD.HYPERLIQUID perpetual strategy that took a directional position only when 4-hour and daily momentum agreed (flat on disagreement) and let winners run on an ATR trailing stop, generalizing the promoted ETH funding-confirmed trend structure by swapping its funding leg for a slower-timeframe momentum leg.

Over 803 days and 173 trades (80 long / 93 short, 3d 16h average hold, 79% exposure) it returned -5.17% with Sharpe -0.047, profit factor 0.923, win rate 35.3%, max drawdown 12.27% and avg_trade_return_pct of -0.109% against the +0.15% venue floor. Gross price PnL before commission was only +$1,136 versus $3,741 of commission and $2,571 of funding paid.

Backtest and paper results are hypothetical. Trading involves risk of loss.