SOL Hyperliquid 12H Momentum-Confluence Trend — Long-Short, Single-Bar Fast/Slow EMA Aligned with a Long Trend EMA, ATR Trailing Exit (Volatility-Scaled, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price MOMENTUM-CONFLUENCE trend follower on SOL-USD perpetual on HYPERLIQUID (SOLUSD.HYPERLIQUID), built on 12H bars. This is a DELIBERATE, evidence-driven fix of the two ways the proven dual-TF mechanism has died on Hyperliquid in my own recent proposals: (1) the LINK-HL 4H version was abandoned as 'promising shape, UNESTABLISHED edge, wrong venue/data' — its edge was genuinely distributed (kurtosis 7.68, skew 0.22, positive every year) but the ~2.3y of 4H HL history was too short to clear the bootstrap-CI / deflated-Sharpe significance gate; (2) the SOL-HL 4H version died fee_edge because 4H trend legs were too small per trade. Moving the SAME proven confluence mechanism to 12H bars fixes BOTH at once: 12H gives ~6+ years of usable Hyperliquid history (≈5000×12H candles → statistically establishable, populated walk-forward + non-empty 15-day holdout), and 12H trend legs are multi-day to multi-week, so per-trade capture clears the ~0.09% round-trip with wide margin. To further cut the verification-loop risk that killed the weekly ETH and multi-bar variants, the signal uses a SINGLE bar type (no multi-timeframe wiring): fast, slow, and long 'trend' EMAs are ALL computed from the 12H series. It keeps the survivor's core logic — align a fast momentum read with a slower trend and go flat on disagreement, trail winners — which is what produces DISTRIBUTED (non-outlier-mirage) edge, unlike the abandoned Donchian breakouts. NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54 — this is a persistent-trend EMA-confluence, not a compression pop), NOT a cointegration/basis pair (dead), NOT a cross-sectional rank basket (L52), NOT a non-price-feed gate (L46), NOT options/COIN-M (L50/L51). Fills the two largest gaps: HYPERLIQUID venue (6.7% vs 20% target) and long-short direction (13.7% vs an 86% long-only book), on the most liquid non-BTC/ETH HL major with large percentage trends. Risk profile: ~1.5% equity risk per trade via ATR-based stop; volatility-scaled sizing (notional = risk_budget / (atr_mult * ATR_12H)) with a hard per-trade notional cap (≤25% of equity*leverage) so no single 12H bar can blow up equity; 2x leverage cap (reads self.config.leverage). Exactly 3 tunable parameters (slow/fast EMA ratio, long trend-EMA length, ATR trailing multiple); fast EMA fixed at the survivor's proven relative value to avoid best-of-N overfit.

Implements the hypothesis exactly: the survivor's confluence mechanism (fast momentum must agree with a slower trend, flat on disagreement, ATR trail runs winners) on SOL-USD Hyperliquid 12H bars, with a SINGLE bar type so there is no multi-timeframe/extra-leg wiring to mis-configure (the verification-loop risk the hypothesis wanted removed) -- no extra_instruments, no extra_bar_types. Exactly 3 tunables (ema_ratio collapsing the fast/slow pair, trend_ema, trail_atr_mult) with the fast EMA fixed at 12; code clamps mirror _param_bounds verbatim. Long-short fills the direction gap, Hyperliquid the venue gap; leverage 2.0 is genuinely consumed via the notional cap (equity * max_notional_frac * self.config.leverage), and the cap is a fraction of CURRENT equity so losses cannot compound into larger positions (L15). VERIFIED BEFORE SUBMITTING by replaying THIS EXACT CLASS over the real catalog series (2475 12H bars): 134 round trips (~40/yr), avg +1.46% per trade net of Hyperliquid's ~0.09% round-trip taker cost -- roughly 16x the fee floor, which is the fee_edge failure the 4H SOL version died of and this timeframe genuinely fixes -- 31% win rate with the trail producing much larger winners than losers, profit factor 1.45, trade-return kurtosis 8.9, top-3 trades 26% of gross profit, and +0.99%/trade over the trailing 365 days. Every parameter setting inside the declared bounds is profitable full-sample (avg +1.01% to +2.19%, PF 1.30-1.73), and the chosen defaults are the cell with the LOWEST tail concentration rather than the highest return, deliberately, to avoid best-of-N overfit. TWO HONEST CAVEATS the reviewer should weigh: (1) the Hyperliquid 12H catalog for SOL starts 2023-01-01 (2475 bars, ~3.4 years) and currently ends 2026-05-22, not the ~6 years / ~5000 candles the hypothesis assumed, so the sandbox will trigger on-demand collection to top up and the significance gate has less data than hoped; (2) profitability is concentrated in 2023 (+4.04%/trade) and 2026 YTD (+1.30%), with 2024 (+0.19%) and 2025 (+0.15%) close to flat, and at ~40 trades/yr a 15-day holdout expects only ~1.6 trades -- so the edge is real but thinly sampled, and a longer-history venue would test the same mechanism more decisively. Per-bar work is O(1) (three incremental EMAs, a 14-deep TR deque), so the 300s Layer-3 cap is safe; all six Layer-2 synthetic scenarios pass locally with a continuous, varying signal.

The 12H 'more data' fix did not deliver, and the edge is worse than the abandoned 4H siblings. The hypothesis claimed 12H would give ~6 years of Hyperliquid history to clear the significance gate, but SOL-HL 12H data starts 2023-01 — only ~3.4 years (2475 bars), the same short single-regime window as the LINK-HL/AVAX-HL 4H versions. The edge is not statistically established (Sharpe 0.56 with CI [-0.35, 1.34] straddling zero) and, worse, not persistent: 2023 (+23.4%) carries essentially the entire result while 2024 (+3.0%), 2025 (+0.2%) and 2026 (+1.5%) are near-flat, with rolling_sharpe decaying from ~4-5 in early 2024 to ~0.6 — the recent window the walk-forward OOS and holdout evaluate has no edge. benchmark_meaningful is true and information_ratio is -0.89 with alpha only 0.046, so it underperforms simply holding SOL. It is the 0/97 Hyperliquid directional-trend class (L41), and the near-identical LINK-HL (Sharpe 0.46, CI-low -0.60) and AVAX-HL (0.51, CI-low -0.57) were just abandoned for this profile; this one is weaker (edge concentrated in 2023, decayed since, IR -0.89). The distributed shape (kurtosis 4.49) is real but a distributed edge that only worked in 2023 is not deployable, and best-of-N over ~225 trials on a Sharpe-0.56 CI-straddling-zero base will not clear deflated Sharpe. Abandon rather than spend the optimization budget.