AvaxDualTimeframeMomentumConfluenceLS

AVAX Dual-Timeframe Momentum Confluence, Long-Short (Single-Instrument AVAXUSDT.BINANCE USD-M, 4H Primary Aligned With 1D Trend, Flat on Disagreement, ATR-Trailing Exit, Clean 3-Parameter — NO Extra Regime/Volume Gate)

A LONG-SHORT, single-instrument, pure-OHLCV trend follower on AVAXUSDT.BINANCE USD-M perpetual. It is a deliberate, evidence-driven re-use of the ONE mechanism family that repeatedly survives this factory's 3-phase deflation and reaches paper: plain DUAL-TIMEFRAME MOMENTUM CONFLUENCE (BTC, ADA, and BNB variants are all currently in paper_stage). It is explicitly NOT the volume-confirmed variant, which the meta-learner (L91) showed does not generalize past ETH, and it is explicitly NOT a regime/efficiency/volatility-gated momentum design (L92) — no extra conditional gate is layered on. The signal is: take a directional position ONLY when a FAST 4H momentum sign agrees with the SLOW 1D trend sign; stand flat when they disagree. AVAX is chosen because it is a liquid, multi-year (2020-listed) Binance USD-M major with strong persistent directional trends, near-zero share of the recent research corpus (which is 62% BTC), and only its volume-CONVICTION-GATED cousin was ever tried here (died overfit — precisely from the extra gate, per L92, not from the base trend edge). Direction is long+short, filling the portfolio's most under-represented direction bucket (85.6% long-only vs a 55% cap).

Implements the hypothesis literally: single instrument (AVAXUSDT.BINANCE USD-M), 4H primary bars for the fast momentum leg and real 1-DAY bars of the SAME instrument as the extra feed for the slow trend leg (so the '1D trend' is an actual daily series, not a proxy EMA), flat on disagreement, ATR-trailing exit, and exactly three tunables (ema_ratio, daily_lookback, trail_atr_mult) with clamps mirrored verbatim in _param_bounds. Structure follows the paper-stage ADA/BTC/BNB confluence survivors rather than the volume-gated (L91) or regime-gated (L92) variants that failed to generalize. Futures venue is correct because the strategy goes short and consumes leverage 2.0 in position_size via self.config.leverage. Entry frequency is deliberately loose (two sign conditions only), so trades fire many times per year on a 4H series; the daily leg is a slower timeframe than the primary so the base template's same-timeframe sync barrier never gates it, and extra_instruments carries AVAXUSDT.BINANCE positionally alongside the daily bar type so the feed resolves. All indicator state is incremental (EMAs, deque-bounded ATR) — O(1) per bar, no history rescans, so the 300s sandbox budget is safe.

analyzing_stuck_3_timeouts

Long-short dual-timeframe momentum confluence on AVAXUSDT Binance USD-M perpetual. A fast 4H EMA-spread momentum read, normalized by ATR (continuous every bar), must agree in sign with a slow 1-DAY trend read (sign of the daily close-to-close return over daily_lookback days). Agreement long -> BUY, agreement short -> SELL, disagreement -> flat. No volume-conviction gate and no regime/efficiency filter is layered on. Exits are (a) an ATR trailing stop that closes when price gives back trail_atr_mult x entry-ATR from the best close since entry, (b) the daily trend flipping against the position, or (c) the 4H spread flipping against the position. Sizing is risk-first: equity * risk_pct / (trail_atr_mult * ATR), capped at max_notional_frac * equity * leverage (2x), so per-trade risk stays ~1.5% of current equity.

Prefer reconstructing side from cache.positions_open(instrument_id=...) entry side rather than the momentary signal sign.

should_exit() infers _side from the sign of the live signal when state is lost (_side==0, e.g. after a restart). If a position is actually open but state was cleared, this could pick the wrong side and mis-evaluate the trailing/flip exits for one bar. In backtest this path is not exercised (state persists), so it is not a correctness bug here; it is a live-robustness note.

Initialize trailing state on fill confirmation (on_position_opened) rather than in should_enter to avoid stale trailing anchors on rejected entries.

_side is set inside should_enter() before order confirmation, and _extreme/_entry_atr are seeded there. If an entry is rejected by the base min-notional check, these fields are set while flat; the next flat-bar should_enter recomputes them, so it self-corrects, but the transient state is slightly leaky.

Sandbox profit is heavily concentrated in the stressed volatility tercile (return +33.6%, Sharpe 1.05) while the calm tercile is nearly flat (+1.3%, Sharpe 0.12) and the FTX-collapse window is negative (-6.9%). Headline Sharpe 0.406 with CI [-0.44, 1.21] straddles zero. Trend-following edge here looks regime-dependent (needs sustained directional AVAX moves); worth checking the calm-regime and recent-year (2024-2026) cells hold up on full history before optimizing.

Trades faithfully implement the stated mechanism: 422 signaled = 422 submitted, zero drops (entry_diag), balanced long 206 / short 216, ATR-trailing exit, confluence gate with flat-on-disagreement. Fair test, not an artifact or mechanism mismatch.

Numerically viable and NOT fee-fragile: avg_trade_return_pct 1.14% is well above the 0.15% Binance USD-M floor, commission is only 3.39% of gross, PF 1.119 (>1.05), positive Sharpe 0.409, DD 32.4% (<60%), 422-trade / 6.7-year reliable sample.

This is the ONE OHLCV family that has reached paper here (Binance USD-M dual-TF momentum confluence: BTC/ADA/BNB siblings in paper_stage), and it is the plain variant that deliberately avoids the volume/regime gates that killed its cousins (L82/L91/L92). Fills the under-represented short direction bucket.

Positive in 4 of 6 full calendar years and across normal+stressed vol terciles; low impact cost (11.6% of gross) and ample capacity ($7.4M).

Edge is mediocre and regime-dependent: Sharpe 0.409 with 95% CI [-0.43, 1.20] straddling zero, PF only 1.119, and profit is concentrated in the stressed tercile (+33.6%, Sharpe 1.05) while calm is nearly flat (+1.3%, Sharpe 0.12) — QA's concern is borne out.

Recent-year deterioration: 2025 +3.9% and 2026 -26.9% (max_drawdown 32.4%, 330-day underwater stretch). If optimization only fits the 2021/2023 bull windows this will fail the holdout/deflation gates.

Sits far below the ADA survivor tier (PF 2.19 / Sharpe 1.60 / DD 10.7%), so the burden is on optimization to find a genuinely robust parameter region rather than curve-fit the two good years.

AvaxDualTimeframeMomentumConfluenceL-5eb1172c2c

This strategy re-used the dual-timeframe momentum-confluence recipe that has repeatedly reached paper here (BTC/ADA/BNB siblings), deliberately dropping the volume and regime gates that killed earlier cousins, and applied it long-short to AVAX to fill an under-represented short-direction bucket. It passed the backtest-review gate with a modest but fee-viable edge (Sharpe 0.41, avg trade 1.14%) and was sent to optimization, which lifted Sharpe to 0.65 and passed the holdout while flagging weak deflation and a failed programme-level FDR. After a single iteration, however, the pipeline abandoned it at the ANALYZING stage due to three consecutive analyst timeouts, so it never received a formal promote/iterate/abandon verdict. It ends recorded as abandoned, with later paper and live stages never reached.

Even a mechanism from the factory's one paper-reaching family can be lost to an operational timeout rather than merit, so the analyzing stage needs resilience to retries; on the merits, this variant's edge was mediocre and regime-dependent (profit concentrated in the stressed vol tercile, weak deflated/FDR statistics) and would likely have struggled at the analyst gate anyway.

It cleared the pre-optimization backtest-review gate (verdict: optimize) and completed the 3-phase optimization, but the run was abandoned at the ANALYZING stage before any analyst promote/iterate/abandon verdict was reached — the abandon reason was a process failure (analyzing_stuck_3_timeouts), not a substantive rejection.

A long-short, pure-OHLCV dual-timeframe momentum-confluence trend follower on AVAXUSDT.BINANCE USD-M that takes a directional position only when a fast 4H EMA-spread sign agrees with the slow 1D trend sign, stays flat on disagreement, and trails winners out with an ATR stop.

The initial backtest over ~6.7 years returned 48.2% total (CAGR 5.72%, Sharpe 0.41 with 95% CI [-0.43, 1.20], PF 1.12, 32.4% max drawdown, 422 trades, avg per-trade return 1.14% of notional). Optimization improved it to 94.9% total return, Sharpe 0.65, PF 1.24 and 23.3% drawdown over 480 trades, and it passed the holdout (Sharpe 0.72, ratio 0.91), though it failed the programme-level FDR (candidate p 0.90, deflated Sharpe ~0.06).