AtomBybitDualTimeframeMomentumConfluenceLS

ATOM Bybit Dual-Timeframe Momentum Confluence, Long-Short (ATOMUSDT.BYBIT 4H Primary + 1D Confirm, Flat on Disagreement, ATR-Trailing Exit, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV momentum-confluence trend follower on ATOMUSDT.BYBIT linear perpetual, 4H primary bars gated by the 1D trend. This deliberately reproduces the ONLY archetype that has reached paper stage in this factory (BTC / ADA Dual-Timeframe Momentum Confluence, both live at Sharpe ~1.6-2.0) but moves it to a venue with ZERO history in the entire 3002-experiment corpus (BYBIT = 0.0%, quota target >=5%) and to a fresh symbol (Cosmos/ATOM) that has never been tried in the momentum family (tried: ADA/BTC/ETH/AVAX/BNB/NEAR/SOL/DOT/XRP/LINK/LTC). It is NOT the banned volume-confirmed variant (L91) and NOT a regime/efficiency gate bolted onto a sub-fee edge (L92): the confirmation leg is a second TIMEFRAME's trend sign, which is exactly the mechanism the two paper-stage winners use, not a volatility/efficiency filter that only thins trades. Rationale for the venue: BYBIT is entirely absent from the corpus yet has clean ~4.9y (2021-10 -> 2026-09) of 4H and 1D ATOM klines in the catalog, taker fee 0.055% (RT ~0.11%), so this is a genuine venue-diversity fill, not another Binance re-test. Only 3 parameters to resist the overfit deaths that killed the 4-parameter variance-ratio and regime-switch clones.

Direct port of the factory's only paper-stage archetype (BTC / ADA Dual-Timeframe Momentum Confluence) onto the venue and symbol the hypothesis targets. The confirmation leg is a second TIMEFRAME's trend sign -- exactly the mechanism the paper-stage winners use -- not a volume gate (L91) and not a volatility/efficiency filter bolted onto a sub-fee edge (L92). Venue is BYBIT per the hypothesis's diversity rationale, and it is the correct routing independently: the strategy takes shorts and requests leverage > 1, so a futures MARGIN venue is required; ATOMUSDT.BYBIT has 4H bars from 2021-10-11 and 1D bars from 2021-10-12 through 2026-09 in the catalog, so both legs have ~4.9y of real data and the sandbox window is fully covered. Entry frequency: sign agreement between an EMA spread and a daily trend flips many times a year on 4H bars, so the joint condition is not the rare-conjunction pattern that produces zero-trade sandboxes (L25/L21); warmup is 73 4H bars (~12 days) plus 19 daily bars, a small fraction of the window. Per-trade edge: exits are trend-following with an ATR trail rather than a fixed few-tenths-of-a-percent target, so the expected per-trade move is multiple ATRs (ATOM 4H ATR is ~1.5-2.5% of price), comfortably above Bybit's ~0.11% round-trip taker cost. Risk-first sizing with a hard notional cap addresses L15 (no equity-compounding blowup) -- the notional never exceeds 0.5x equity. extra_instruments carries ATOMUSDT.BYBIT positionally alongside the daily bar type so the resolver actually attaches the slow feed; on_extra_bar dedupes by timestamp because the daily lookback indexes by position.

Cost-fragile, no deployable edge on a decisive 327-trade / 4.9-year sample. profit_factor 1.0218 sits in the fee/impact-drag band [0.85,1.10] (L18) and the dominant cost is market IMPACT, not commission: impact_cost_pct 61.2% consumes ~two-thirds of gross PnL and capacity_usd is only $266k, so the nominal 0.296% per-trade return survives only at toy scale and vanishes at any meaningful size (QA's port-to-alt concern borne out). Sharpe 0.0685 with CI [-0.842, 0.996] straddling zero, Sortino 0.084, recovery_factor 0.18. The book LOSES in both the calm (-9.0%) and stressed (-7.6%) vol terciles with all PnL from the 'normal' tercile, and loses badly in 2025 (-12.8%). This is the single-name 4H dual-TF momentum-confluence clone (L82/L93) sitting far below the only surviving tier (Binance-USD-M ADA at Sharpe 1.60/PF 2.19); the identical ATOM confluence was abandoned yesterday (ATOM-BTC macro, impact 84.6%, capacity $139k). A PF-1.02 signal whose gross edge is already 61% consumed by impact cannot be tuned above costs by 3 parameters — not worth 2 hours of optimization. abandon_class=negative_expectancy.

Long-short dual-timeframe momentum confluence trend follower on ATOMUSDT.BYBIT linear perpetual. The fast leg is a volatility-scaled EMA spread on 4H bars, (EMA_12 - EMA_34)/ATR_14, computed incrementally every bar and returned as the continuous signal. The slow leg is the sign of the 18-day return read from an extra 1-DAY ATOM feed in on_extra_bar. Entry requires CONFLUENCE: 4H spread > 0 and daily direction > 0 goes long, spread < 0 and daily direction < 0 goes short, any disagreement (including a neutral daily read) stays flat. Exits on whichever comes first: an ATR trailing stop giving back trail_atr_mult x entry-ATR from the best close since entry, the daily trend flipping against the position, or the 4H spread flipping against the position. Sizing is risk-first -- equity * risk_pct / (trail_atr_mult * ATR) so the trailing-stop distance risks ~1.3% of equity per trade -- then capped at max_notional_frac * equity * config.leverage (0.25 * 2x = 0.5x equity notional), which is where the requested 2x leverage is actually consumed. Only three tunables (ema_ratio, daily_lookback, trail_atr_mult), each clamped to a range declared in _param_bounds.

On live restart, reconcile self._side from the actual open position in cache.positions_open() rather than from the signal sign.

should_exit() infers self._side from the live signal when state is lost (self._side == 0). This is a reasonable restart-recovery guard, but in a restart scenario it could infer the wrong side if the open position and the current signal disagree, delaying the correct exit. Backtests run in one process so this path is never hit; noting for live-parity awareness only.

Confirm the base template calls the override for warmup gating (it does) — no change required, deliberate override.

min_bars_required() overrides the framework method (static analysis base_shadow warning). The override (~73 bars) is deliberate and larger warmup only; sandbox produced 327 trades so warmup does not starve entries.

Sandbox (smoke test, unoptimized) shows Sharpe 0.068 and a regime split that LOSES in both the calm (-8.96%, Sharpe -0.20) and stressed (-7.57%, Sharpe -0.23) vol terciles, with all positive PnL concentrated in the 'normal' tercile (+22.76%). This is the single-name momentum-confluence archetype that has repeatedly died on full history when ported to alts (LTC/SOL/ETH-BTC etc., avg_trade_return_pct going negative). The port-to-alt premise (ATOM behaves like BTC/ADA) is the open question for the analyst — the mechanism may not survive full-history net of costs. Advisory only; correctness is fine, so it proceeds to a real backtest where performance is judged.

Reliable, decisive sample: 327 trades over 1791 days / ~4.9y of clean 4H+1D ATOM klines (metrics_reliable=true).

Code is a faithful implementation of the dual-TF confluence mechanism (327 signaled = 327 submitted, balanced long/short 153/174, ATR-trailed exit, daily-trend gate) — this is a fair test, not a broken build.

Drawdown is contained (17.99%) and nominal avg_trade_return_pct 0.296% is above the ~0.15% BYBIT floor.

profit_factor 1.0218 sits in the fee/impact-drag band [0.85,1.10] — gross edge barely exceeds costs.

impact_cost_pct 61.2% consumes ~two-thirds of gross PnL and capacity_usd is only $266k, so on thin ATOM the apparent edge survives only at toy scale and evaporates at any deployable size.

Sharpe 0.0685 with CI [-0.842, 0.996] straddling zero; Sortino 0.084, recovery_factor 0.18, Calmar 0.037 — no risk-adjusted edge.

Single-name 4H dual-TF momentum confluence clone (L82/L93): the only survivor of this family is Binance-USD-M ADA (Sharpe 1.60/PF 2.19/DD 10.7%); this BYBIT ATOM port is Sharpe 0.07/PF 1.02, far below that tier, and the identical ATOM confluence was abandoned yesterday (ATOM-BTC macro, impact 84.6%, capacity $139k).

QA's port-to-alt concern confirmed on full history: loses in the calm (-9.0%) and stressed (-7.6%) vol terciles, all positive PnL confined to the 'normal' tercile; loses in 2 of the last 2 recent years (2025 -12.8%).

AtomBybitDualTimeframeMomentumConflu-eb34d5f4a5

This strategy cloned the factory's only paper-stage winner (dual-timeframe 4H/1D momentum confluence) onto BYBIT ATOM to fill a venue-diversity gap, keeping to just three parameters to avoid overfit deaths. On a clean, decisive 327-trade / 4.9-year sample the mechanism worked as coded (balanced 153 long / 174 short, contained 18% drawdown) but produced no real edge: Sharpe 0.069, profit factor 1.02, and market impact devoured 61% of gross PnL against a thin $266k capacity. The analyst abandoned it at the pre-optimization backtest-review gate, noting it sat far below the surviving Binance-USD-M ADA tier (Sharpe 1.60/PF 2.19) and that an identical ATOM confluence had been abandoned the day before. It never reached optimization, post-optimization analysis, or risk review.

Faithfully replicating a winning archetype's mechanism does not transfer edge to a new low-liquidity venue/symbol -- a signal whose gross edge is already ~61% consumed by impact at $266k capacity cannot be tuned above costs by three parameters.

The backtest-review analyst returned an 'abandon' verdict before optimization: no deployable risk-adjusted edge (Sharpe straddling zero, PF 1.02 in the fee/impact-drag band) with two-thirds of gross PnL eaten by market impact at toy scale, plus losses in both the calm (-9.0%) and stressed (-7.6%) vol terciles and in 2025 (-12.8%).

Ported the factory's only paper-stage archetype -- a 4H/1D dual-timeframe momentum-confluence trend follower (long when both timeframes agree up, short when both agree down, flat on disagreement, ATR-trailing exit) -- onto ATOMUSDT on BYBIT, a venue with zero corpus history and a symbol never tried in the momentum family.

Over a decisive 327-trade, ~4.9-year sample (2021-10 to 2026-09) it returned 7.38x total (CAGR 67.5%) but with a Sharpe of only 0.069 (CI [-0.842, 0.996]), profit factor 1.02, win rate 38.5%, and max drawdown 18.0%. Impact cost consumed 61.2% of gross PnL and capacity was only $266k, with per-trade return 0.296% barely above the ~0.15% floor.