AvaxConvictionGatedMomentumLS

AVAX Conviction-Gated Momentum, Long-Short (AVAXUSDT.BINANCE, 4H — STRONGER filter: volume surge AND directional close-location)

Implements the hypothesis's stronger conviction gate without turning it into a hard AND-conjunction that would fire too rarely (a known zero-trade failure mode): volume surge and directional close location enter as multiplicative, continuous conviction factors on the momentum z-score, so the signal varies every bar while only crossing the entry threshold when momentum is strong AND the bar was volume-backed AND closed toward the direction of the move. Long+short with no leverage requires a MARGIN venue, so BINANCE USD-M is kept as specified (leverage 1.0, referenced in the notional cap). An offline replay of the exact signal on the catalog's 13,037 AVAXUSDT 4H bars gives ~530-600 trades over ~6 years (~90-100/yr, avg hold ~12 bars) with positive per-trade expectancy net of the ~0.10% round-trip taker cost across the entry-threshold 0.9-1.2 and stop/TP neighbourhood, so the design is not sitting on a parameter cliff and clears the fee floor.

Overfit / no-generalization death, NOT a fee-edge death (per-trade return 0.53% clears the floor). Two unwaivable HARD gates fail: walk-forward is_overfitted=TRUE (IS Sharpe 1.218 -> OOS 0.091, with per-window OOS -0.18/-0.72/+1.18, i.e. two of three negative) and PBO 0.6693 > 0.5 (selection more likely than not overfitting). Robustness confirms the selected Sharpe is best-of-N noise over 225 trials: deflated_sharpe 0.0072 << 0.95, is_significant=false, Sharpe CI [-0.47, 1.11] straddles zero, and it does not survive programme-level FDR (keeps 3 of 238, candidate_p 0.993). Sensitivity flagged 7 cliff parameters, so there is no robust parameter region to tune toward — iteration can only curve-fit the noise again. Per the overfit-vs-dead-premise rule, high-IS/OOS-collapse is ABANDON not revise: this AVAX single-name momentum family has already died repeatedly (AVAX dual-TF momentum-confluence -> fee_edge; AVAX daily dual-EMA trend -> overfit x3 iters; AVAX funding-gated momentum -> overfit; AVAX Keltner breakout -> overfit), so re-pointing the same conviction-gated momentum recipe at another horizon/instrument just overfits again. This is the single-name trend/momentum clone trap (L90): below the ADA-winner tier, the base backtest (Sharpe 0.31, PF 1.12) is the ceiling and deflation only lowers it. Failure pattern: overfit / no-generalization.

AVAXUSDT 4H long-short momentum on Binance USD-M futures. Each bar computes a volatility-normalised 12-bar momentum z-score and scales it by two continuous conviction factors: a volume-surge factor (bar volume / 20-bar mean volume, divided by a 1.3x surge reference, capped at 2.5) and a directional close-location factor (how far the close sits toward the bar extreme that agrees with the momentum sign). The product is the signal; |signal| > 1.0 opens a long (positive) or short (negative). Exits are venue-managed bracket legs (5% stop / 12% take-profit, firing intra-bar) plus signal exits: momentum decay/flip (z crosses back through +-0.25 against the position) or a 30-bar (5-day) time stop. Sizing is risk-based: notional = equity * 1.2% / 5% stop (~0.24x equity), capped at 60% of equity.

If the base relies on its own warmup accounting, ensure the override is honored or fold the requirement into the documented parameter path.

min_bars_required() overrides a base framework method; static analysis flagged base_shadow. This appears deliberate (used to set the warmup) and the value matches the deque sizing, so it is not a defect, but confirm the base template actually consumes this override for warmup gating rather than its own internal default.

Sandbox is only a marginal, positive-but-weak result: Sharpe 0.25 (CI [-0.53, 1.09] straddles zero), win_rate 0.35, profit_factor 1.10, information_ratio -0.58 vs buy-hold, and max_drawdown 37%. avg_trade_return_pct 0.44% clears the fee floor and the edge is spread across regimes (calm/normal/stressed all positive), so the mechanism is not obviously broken, but the risk-adjusted edge is thin and the optimizer will need to find a genuinely robust parameter region rather than curve-fit the 565-trade sample. Worth watching the walk-forward OOS closely; conviction-gated momentum on a single alt at 4H has a real overfit risk.

Reliable sample: 565 trades over 6.0 years (metrics_reliable=true) — enough for meaningful sensitivity/walk-forward analysis.

avg_trade_return_pct 0.44% clears the 0.15% Binance USD-M fee floor comfortably; commission is only 3.41% of gross, so the edge sits ABOVE trading costs (not a fee-edge death).

Trades match the hypothesized mechanism: balanced 302 long / 263 short conviction-gated momentum, 31% exposure (not an always-in beta ride), positive in all three vol regimes.

No artifact signature: total_return 41% is realized (end_unrealized_pct 7.7%), Sharpe from mtm equity curve, plausible not implausible.

Thin risk-adjusted edge: Sharpe 0.25 with CI [-0.58, 1.07] straddling zero, profit_factor 1.08, win_rate 0.345.

Negative information_ratio (-0.58) vs buy-hold and 36.8% max drawdown; losing years in 2020/2021 and 2026 YTD.

Single-alt 4H conviction-gated momentum carries real overfit risk (QA edge concern) — the optimizer must find a genuinely robust parameter region rather than curve-fit the 565-trade sample.

Reliable 556-trade / 6-year sample; avg_trade_return_pct 0.53% clears the 0.15% Binance USD-M fee floor, so this is NOT a fee-edge death (commission only 3.53% of gross).

Holdout is technically positive (Sharpe 0.78 on 88 trades) and optimization did reduce drawdown (36.8% -> 23.1%) and lift Sortino.

Walk-forward is_overfitted=TRUE: in-sample Sharpe 1.218 collapses to out-of-sample 0.091, with per-window OOS Sharpes -0.183 / -0.722 / +1.177 (two of three negative) — the edge does not generalize, it fit noise.

PBO 0.6693 > 0.5 — the parameter selection is more likely than not overfitting (HARD gate).

Deflated Sharpe 0.0072 << 0.95, is_significant=false, Sharpe CI [-0.47, 1.11] straddles zero, and it does NOT survive programme-level FDR (keeps 3 of 238) — the selected Sharpe is indistinguishable from best-of-N noise over 225 trials.

Sensitivity found 7 cliff parameters (volume_ma, vol_window, entry_thresh, mom_lookback, stop_loss_pct, vol_surge_mult, take_profit_pct) — the config sits on a knife-edge, the opposite of a robust region.

Pre-registered prediction met only 2 of 6 (missed OOS Sharpe 1.1, PF 1.35, win_rate 0.42, DD 20%).

Holdout 'pass' rides the overfit params and its ratio 8.58 vs WF-OOS 0.09 is only high because the WF-OOS is near zero, not because the holdout is strong.

Benjamini-Hochberg at q=0.10 over 238 programme candidates keeps 3. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

AvaxConvictionGatedMomentumLS-3570323b98

AvaxConvictionGatedMomentumLS gated 4H AVAX momentum by volume surge and directional close-location, taking balanced long and short positions. Its base backtest was viable on cost grounds — 0.53% avg per-trade return over 565 trades — so the reviewer sent it to optimization, but the numbers were thin: Sharpe 0.25 rising only to 0.31, with a confidence interval straddling zero. Optimization exposed the strategy as overfit: in-sample Sharpe of 1.22 collapsed to 0.09 out-of-sample, PBO was 0.67, deflated Sharpe was 0.007, and seven parameters sat on cliffs. The analyst abandoned it as an overfit / no-generalization death — part of a repeatedly-failing single-name AVAX momentum family where the base backtest is the ceiling and iteration only curve-fits noise again.

A per-trade edge above the fee floor is not enough — when in-sample Sharpe collapses out-of-sample and the config sits on parameter cliffs, the edge is best-of-N noise, and re-pointing the same single-name momentum recipe at a new horizon only overfits again.

The analyst abandoned it after optimization on two unwaivable HARD gates: walk-forward is_overfitted=TRUE (in-sample Sharpe 1.218 collapsing to OOS 0.091, two of three windows negative) and PBO 0.669 > 0.5, with deflated Sharpe 0.0072, an insignificant result that fails programme-level FDR, and 7 sensitivity cliff parameters.

A long-short 4H AVAXUSDT momentum strategy that scaled a normalized momentum z-score by two continuous conviction gates — a volume-surge factor and a directional close-location factor — entering only when momentum was strong, volume-backed, and closed in the direction of the move.

The base backtest over 6 years (565 trades, 302 long / 263 short) returned 41.1% with Sharpe 0.25 (CI [-0.58, 1.07]), profit factor 1.08, win rate 34.5%, and 36.8% max drawdown; optimization improved it modestly to Sharpe 0.31, profit factor 1.12 and 23.1% drawdown, with avg_trade_return_pct 0.53% clearing the 0.15% fee floor.