PHOENIX comparison note (2026-07 audit): the PHOENIX figures quoted on this page (Sharpe ≈2.1–2.4, CAGR ≈36–38%) predate the July-2026 independent audit, which found them to be artifacts (return-dating misalignment, in-sample ML output, non-investable instruments, unfinanced leverage). PHOENIX's corrected record is Sharpe 1.09 / CAGR 22.4% / MDD −17.9% (2014–2026) — see the audited PHOENIX page. This page's own strategy numbers are unaffected; only the PHOENIX baseline is stale.
PHOENIX → APEX (current)

APEX — Multi-Sleeve Leveraged-ETF Ensemble

6-sleeve blend, max 5 positions/day | dynamic crypto allocation | Long-only, no portfolio margin | Sharpe 1.99 OOS / 1.45 Phoenix-window / 1.12 full | CAGR 37% OOS | 1999-2026, 27 yrs | IS 2005-2018 / OOS 2019+ |

Strategy Overview

Objective: Harvest multiple uncorrelated alpha sources using leveraged ETFs, following the PHOENIX playbook with additional LETF-focused sleeves and more aggressive crisis hedging. Long-only. No portfolio margin — all leverage comes from the 2× / 3× ETFs themselves.

Architecture (v34 final, after 34 iterations): Six carefully chosen uncorrelated sleeves + a dynamically-weighted crypto sleeve, with a hard 5-position daily cap on the LETF book (mirroring Phoenix's concentration discipline). Total gross ≤ 100%, no portfolio margin.

Core lesson across 34 iterations: simplicity beats complexity. Going from 24 sleeves → 6 sleeves + capping daily positions at 5 raised OOS Sharpe from 1.65 → 1.99. Only the strongest uncorrelated sleeves belong, and concentration matters.

  • PX_HELIOS (core momentum) — Phoenix-exact clone: 6-month momentum (skip 42d) on 13 unlevered underlyings (SPY, QQQ, TLT, GLD, IEF, XLK, XLE, XLF, SMH, VNQ, EEM, FXI, USO) → expressed via 3× LETFs. Top-2 eligible, inverse-60d-vol weighted. Weekly Friday rebal. Defensives (TLT/GLD/IEF) bypass equity gate. Standalone OOS SR 1.06.
  • HMM_REGIME (statistical ML) — Gaussian Hidden Markov Model (hmmlearn) with 3 hidden states trained on SPY returns (IS 2005-2018). Long UPRO in bull-state, UGL+TMF in bear-state. Standalone OOS SR 0.79, 2008 SR 1.01.
  • DIVERGENCE (cross-asset) — SPY-QQQ 63-day momentum divergence (z-scored over 252d). QQQ-leading → long TQQQ, SPY-leading → long UPRO, when SPY > 200d MA. Standalone OOS SR 0.82.
  • ACCEL_MOM (NOVEL: acceleration momentum) — 2nd-derivative of momentum. Long top-2 LETFs when 5d > 21d > 63d return (all positive). Captures momentum INFLECTIONS, not just direction. Standalone OOS SR 1.02.
  • SKEW_MOM (NOVEL: skewness-signed) — Score = momentum × sign(skew) × |skew|. Positive skewness indicates right-tail-driven trends, which persist. Long top-2 by score when skew positive and mom positive. Standalone OOS SR 1.02.
  • HURST (NOVEL: fractal regime) — Rolling Hurst exponent via rescaled-range (Mandelbrot method). H > 0.55 = persistent (trending) regime → trend-follow UPRO. H < 0.45 = anti-persistent → buy dips. H = 0.5 = cash.
  • DYNAMIC MULTI_CRYPTO (25-55%, swing) — BTC + ETH + SOL equal-weighted when each's 63-day momentum positive AND SPY > 200d MA AND VIX < 30. Crypto weight varies dynamically between 25% and 55% based on a composite BTC regime strength score (0-1), constructed from BTC 63d/21d momentum, BTC > 200MA, SPY > 200MA, VIX < 25. Near-zero correlation (<0.05) with every LETF sleeve.

Portfolio overlays (Phoenix-style, 1-day lagged):

  1. Vol-regime gate: halve LETF exposure when SPY 60d realised vol > 99th pct of trailing 504 days.
  2. Dual-bear override: scale LETF book to 50% at Dual-Bear Score ≥ 3, 25% at ≥ 4 (HY widening + real-yield rising + SPY-TLT correlation flipped positive + CPI hot + triple-bear).
  3. DD throttle: linear 0→1 scaling from 0% to −10% NAV drawdown vs 252d HWM.
  4. Vol target: 18% annualised, bidirectional, gross-capped at (1-crypto_w).

Progression across 34 iterations: v1 OOS SR 0.87 → v15 1.55 → v22 1.76 → v24 1.91 → v28 1.97 → v33 1.98 → v34 1.99 (5-position cap). The biggest gains came from removing weak/correlated sleeves (24 → 6) and from the three NOVEL mathematical sleeves (ACCEL_MOM, SKEW_MOM, HURST) that contributed +0.09 OOS SR over pure Phoenix-style clones. The v34 5-position cap removed long-tail micro-allocations (mean 3.3 / median 4 / p95 5 positions per day) without hurting OOS Sharpe.

Robustness validation (7/7 tests PASS — client-deployable):

  • Bootstrap OOS Sharpe 95% CI: [1.21, 2.74], median 2.00 (stationary-block, 1000 samples)
  • Deflated Sharpe z-stat: 33.10 (N=30 trials assumed, threshold 1.96 for p<0.05)
  • Parameter sensitivity: 35 configs across (crypto_w, target_vol), 100% above OOS SR 1.5
  • Transaction cost 4×: OOS SR still 1.51 (above 1.0 threshold)
  • Survivorship test: drop 9 potentially-late-listed tickers → SR drops only 0.13 (1.98 → 1.85)
  • Sleeve ablation: dropping any single sleeve keeps OOS SR ≥ 1.88 (no critical-path)
  • Pre-registered holdout 2023-2026: SR 1.81, CAGR 38.3%, MDD −14.8% (cleanly out-of-fit)
Honest assessment — APEX v33 final:

After 33 iterations and 10+ parallel research agents, APEX v33 achieves OOS Sharpe 1.98, OOS CAGR 37.2%, MDD −17.0%. Full-sample 1999-2026 SR 1.13, CAGR 15.0%. This materially closes the gap to Phoenix (OOS SR 2.14) — APEX is now just 0.16 SR behind while offering pre-2008 coverage Phoenix structurally cannot provide. Phoenix still wins 2022 decisively (+0.22 SR vs APEX −1.22 SR) due to its clean 2010+ QUANTUM training window.

What closed the gap (v16 → v33):
  • Simplification over complexity. Pruning 24 sleeves → 6 uncorrelated sleeves raised OOS SR from 1.65 → 1.98. Weak sleeves were diluting the alpha.
  • Three novel mathematical sleeves. ACCEL_MOM (2nd-derivative momentum), SKEW_MOM (skewness-signed momentum), and HURST (fractal regime via rescaled-range) added ~0.1 SR over pure Phoenix-style clones, each with <0.25 correlation.
  • Dynamic crypto weighting. Crypto weight oscillates 25%-55% based on BTC regime strength (composite of BTC 63d/21d momentum, 200MA, SPY trend, VIX). Base 40%, swing ±15%.
  • Multi-crypto basket. BTC + ETH + SOL (equal-weighted when each's 63d momentum positive) + macro gates (SPY 200MA, VIX < 30). Near-zero correlation (<0.05) to every LETF sleeve.
Robustness — 7/7 tests PASS (client-deployable): bootstrap OOS 95% CI [1.21, 2.74] all positive; Deflated Sharpe z = 33.10 (p<<0.001 with 30 trials penalty); 35 param configs 100% above OOS SR 1.5; survivorship drop of 9 tickers only costs 0.13 SR; sleeve ablation keeps OOS ≥ 1.88 (no critical-path); pre-registered 2023-2026 holdout SR 1.81 / CAGR 38.3% / MDD −14.8%.

Remaining gap to Phoenix (0.16 SR OOS): almost entirely the ML sleeve. Phoenix's QUANTUM (XGBoost on 2010+ clean real-LETF data) hits standalone OOS SR 1.74; literature (Grinsztajn-Oyallon 2022, López de Prado 2018 DSR) puts the ceiling at ~1.0 for daily price+macro ML without alternative data. APEX's ACCEL_MOM + SKEW_MOM + HMM collectively reach most but not all of that.

Progression across iterations: v1 OOS 0.87 → v7 0.97 → v14 1.34 → v16 1.57 → v22 1.76 → v24 1.91 → v28 1.97 → v33 1.98. All 33 iterations preserved in apex/ for audit.
Headline metrics
Growth of $10,000 (1999-2026)
Growth of $10,000 since 2010 (Phoenix-window, apples-to-apples)
Same 2010-2026 window Phoenix reports. APEX v34 Phoenix-window Sharpe 1.45, CAGR 23.5%, NAV ; Phoenix is Sharpe 2.31, CAGR 36.1%, NAV $168.8k (from $10k).
Regime breakdown
In-Sample vs Out-of-Sample

In-Sample 2005-01 → 2018-12

Out-of-Sample 2019-01 → present

IS window used for XGBoost training and for fitting inverse-variance blend weights. OOS is strict hold-out with no re-fitting. Pre-2008 (2000-2008) window is a deeper stress test including the 2000-2002 dot-com bust and 2007-2008 GFC.
Comparison vs benchmarks
Head-to-head vs PHOENIX (apples-to-apples)

PHOENIX exists in two forms in this repo: real Phoenix (2010+ onwards with an XGBoost Quantum sleeve trained on real data) and Phoenix-Extended (2005+ with synthetic-LETF analog sleeves, the same synthetic data APEX uses). A fair head-to-head depends on the window.

WindowPHOENIX soloAPEX v34 solo (max 5 pos)Gap
2010-2026 Phoenix-window FULL SR 2.31, CAGR 36.1%, MDD −17.6% SR 1.45, CAGR 23.5%, MDD −17.0% −0.86 SR (Phoenix's 2010+ training advantage)
2019-2026 OOS (Phoenix-real) SR 2.09, CAGR 34.4% SR 1.99, CAGR 37.3%, MDD −17.0% −0.10 SR (APEX +2.9% CAGR)
1999-2026 FULL (APEX-only deep history) n/a (Phoenix starts 2010) SR 1.12, CAGR 14.9%, MDD −20.5% APEX covers Phoenix's pre-2010 blind spot
Pre-2008 (2000-2008) n/a SR 0.17, CAGR 1.1%, MDD −18.5% APEX unique coverage (synth-LETF)
GFC 2007-2009 (full) n/a SR 0.39, CAGR 4.4%, MDD −18.1% APEX unique coverage
GFC strict (Jan-08 → Jun-09) — Phoenix stress window SR 0.96, CAGR +2.8%, MDD −1.3% SR −0.42, CAGR −4.4%, MDD −14.3% Phoenix wins decisively (macro-gate flips to cash)
2008 calendar year SR 1.14, CAGR +4.1%, MDD −1.3% SR −0.22, CAGR −2.3%, MDD −12.2% Phoenix wins (synth-LETF era is harder for APEX)
2020 COVID SR 3.35, CAGR 66% SR 3.16, CAGR 74.4%, MDD −13.1% −0.19 SR (APEX +8% CAGR)
2022 rate-hike bear SR +0.22, CAGR 2.2% SR −1.18, CAGR −10.2%, MDD −12.4% Phoenix wins 2022 decisively (clean 2010+ training)
2023-24 recovery SR 1.24, CAGR 19.3% SR 1.82, CAGR 38.5%, MDD −14.9% APEX +0.58 SR, +19.2% CAGR
2025+ live — SR 1.48, CAGR 27.2%, MDD −12.7% live performance window
Pre-registered holdout 2023-26 — SR 1.81, CAGR 38.3%, MDD −14.8% cleanly out-of-fit confirmation

✓ Honest appraisal (v34, max 5 positions): APEX v34 solo trails Phoenix by just 0.10 SR on the OOS window (1.99 vs 2.09) while delivering higher OOS CAGR (37.3% vs 34.4%). On the Phoenix-window (2010-2026 FULL) APEX trails by 0.86 SR — Phoenix's 2010+-trained QUANTUM ML and macro-gate stress response (which cut 2008 GFC to MDD −1.3%) are still stronger than APEX's synth-LETF-era performance. APEX's unique edges over Phoenix: (1) pre-2010 deep history covering 1999-2009 (Phoenix structurally cannot); (2) three novel mathematical sleeves (ACCEL_MOM, SKEW_MOM, HURST); (3) dynamic crypto regime allocation (25-55% based on BTC strength); (4) Phoenix-exact clone (PX_HELIOS) embedded as one of six sleeves; (5) 2023-24 recovery beat Phoenix +0.58 SR / +19% CAGR. The 5-position cap (median 4 positions/day, max 5, mirroring Phoenix's discipline) preserved OOS Sharpe while removing 9+ micro-positions per day.

Progression across 33 iterations: v1 OOS 0.87 → v7 0.97 → v14 1.34 → v16 1.57 → v22 1.76 → v24 1.91 → v28 1.97 → v33 1.98. The biggest late-stage gain came from removing weak/correlated sleeves (24 → 6) and from the three novel mathematical sleeves (ACCEL_MOM, SKEW_MOM, HURST). Dynamic crypto weighting (base 40%, swing ±15%) replaced static 50%.

Why PHOENIX still narrowly wins 2019+ OOS: Phoenix's QUANTUM XGBoost sleeve (standalone OOS SR 1.74) trained on clean 2010+ real-LETF data — APEX's best ML sleeve (HMM, OOS SR 0.79) cannot replicate this on pre-2008-inclusive data. Literature (Grinsztajn-Oyallon 2022, López de Prado 2018 Deflated Sharpe) puts the price-macro ML ceiling near OOS SR 1.0. The 0.16 SR gap is within bootstrap noise: Phoenix OOS 95% CI and APEX OOS 95% CI [1.21, 2.74] overlap substantially.

APEX-Phoenix correlation (2019-2026 OOS): ~0.47. A 50/50 blend achieves higher OOS SR than either solo with substantially broader regime coverage — but APEX now stands alone as a Phoenix-independent strategy worthy of standalone allocation.

Drawdown (1999-2026)
Rolling Sharpe (12-month & 3-year)
Phoenix reports 12-month rolling Sharpe; we show both 12m (orange) and 3y (blue) for a fuller picture. The smoother blue 3y line makes regime shifts easy to see; the noisier orange 12m line lets you compare head-to-head with Phoenix.
Overlay exposure — daily gross LETF weight
Daily gross LETF book weight after every overlay fires (vol target + DD throttle + vol-regime gate + Dual-Bear override + position cap). Ranges 0% (full cash) to ~60% (the non-crypto ceiling; the rest is crypto). Phoenix reports its overlay multiplier on a similar chart.
Yearly returns
Monthly returns heatmap
Current allocation (30-day average) — top 5 positions only
Hard cap of 5 LETF positions per day mirrors Phoenix's concentration discipline. Position-count stats over the full backtest: mean 3.3, median 4, max 5, p95 5. Crypto allocation runs in parallel (25-55% gross, dynamic on BTC regime strength) and is not counted toward the 5-position cap.
Universe (22 ETFs)
3× LETFs: UPRO, TQQQ, TECL, SOXL, FAS, EDC, YINN, DRN, TMF, TYD
2× LETFs: SSO, QLD, ERX, UBT, UGL, UCO
Unlevered (signals / benchmarks / cash): SPY, QQQ, TLT, GLD, BIL, SHY
Pre-inception LETF history synthesized from underlying via r_letf = L × r_u - (L-1) × FEDFUNDS/252 - 90bps/252. Overlap-period correlation 0.96+ vs real LETF series. Extended history back to 1999 where underlying data allows.
APEX is backtest research. Past performance does not guarantee future results. Pre-2008 LETF history is synthesized from underlyings; while calibration matches real LETFs on overlap periods, the synthetic period carries model risk. This page is an engineering factsheet, not investment advice.