meta optuna blend

DeMiguel-Garlappi-Uppal (2009, RFS) "Optimal vs Naive Diversification" shows 1/N portfolio weighting is shockingly hard to beat out-of-sample once you account for estimation error in fancy optimization. We apply this at the alpha-family level: K=5 canonical sub-signals (SMA50/200 cross, RSI(14) mean-reversion, 20d-high breakout, 1m-vs-12m TSMOM-lite, 5d return reversion), each returns {-1, 0, +1}, the meta position is the mean clipped to [-1, +1] — buy when ≥3/5 agree long, short when ≥3/5 agree short.

Lopez de Prado 2016 introduces HRP as a quasi-optimal allocator that avoids the unstable matrix-inversion at the core of mean-variance. HRP delivers higher Sharpe out-of-sample than 1/N (DeMiguel-Garlappi-Uppal 2009) and minimum-variance, especially when asset count grows. Recipe: correlation matrix → distance matrix → single-linkage hierarchical clustering → quasi-diagonalization → recursive bisection inverse-variance allocation. Applied here over the K=5 canonical sub-signals (same set as meta_equal_weight) for head-to-head HRP-vs-1/N comparison on every ticker.

Two-lane allocator that combines the Sharpe-weighted top-K from the TA-strategy bench (RSI/MACD/Bollinger/etc., 22 components) with the Sharpe-weighted top-K from the alpha-family bench (91 components, diversity-guarded). Each lane independently produces a blended position series; the allocator combines them with weight w_A (TA) + w_B (alpha) = 1, default 0.50/0.50. Reduces signal-specific failure modes by averaging over heterogeneous information sources at the same execution layer — TA captures price-pattern alpha, alpha-families capture event/fundamental/sentiment alpha, and the ensemble harvests both. Surfaces when neither bench produces a dominant single-strategy champion (cup_tier ≥ 2). On the May 2026 true-struggling cohort (46 tickers, no source cup_tier > 1) the ensemble + per-bench blends lifted cup_tier by an average of +1.43 levels on 65% of (ticker, profile) pairs.

Take this ticker's most recent user_strategies champion, replay its base position from the recorded indicator templates, then apply each Alpha Discovery V2 overlay filter on top: low-vol regime, earnings blackout, low FINRA short-pct, VIX contango, SPY uptrend, sector uptrend. Goal: surface cases where the existing champion would benefit from a regime/event/flow filter that the old finder doesn't apply.