Seven rule-based LETF strategies — VANGUARD · ORION · HELIOS · CRYPTO · REVERSAL · TURN-OF-MONTH · BONDTREND — equal-weighted at 1/7 each (1/6 before CRYPTO activates at IBIT's Jan 2024 listing). No portfolio margin. No borrowing. Nothing fitted. Audited three times; every number on this page is the post-audit v4 series.
Earlier versions of this page advertised a Sharpe above 2 and a 36% CAGR. Those numbers did not survive audit. The reconciliation:
The full audit trail lives in the repository: alt/PHOENIX_REVIEW.md (the v2 teardown), alt/PHOENIX_V3.md (the rebuild), and alt/PHOENIX_V4_REVIEW.md (the re-audit that produced the numbers shown here). The live record starts 2026-07-02; everything before that date is backtest, and backtests are backtests.
A GitHub Actions cron runs the strategy at 6:30pm ET each weekday. Below: explicit BUY/SELL tickets, a share-count calculator for your portfolio size, and the inputs that drove the decision.
A virtual portfolio that fires every live signal at the next-day market open using the same yfinance Open prices the backtest uses. The tracking error against the backtest tells you exactly how much friction live execution will eat.
Scale the Phoenix backtest's net-of-cost return curve to any starting amount over any window. Compare against buy-and-hold SPY and 60/40 with the same data, overlay, and TC drag the live signal uses.
Seven different traders. Seven different signals. One equal-weighted portfolio. One tail-risk brake. No borrowing, no margin, no shorting.
Instead of picking one “best” way to trade, PHOENIX runs seven different trading strategies at the same time and blends them in equal parts. Each uses a different kind of signal — three follow trends in different ways, one buys sharp dips, one trades a calendar effect, one follows the bond market, one follows Bitcoin and Ether. Honesty requires saying the quiet part out loud: the three trend-followers often win and lose together. Most of the smoothing comes from the other four.
Three of the seven are variations of the same trend bet. The diversification is real — but modest, not magic. Why seven, not just one
Blended together, some of the ups and downs cancel and the ride is smoother than most single sleeves — though not dramatically so. On top of that, one simple “brake” halves exposure automatically in the rare weeks when market volatility hits extremes.
Swipe →
Every single one of these strategies has years where it loses money. But an honest caveat first: this page used to claim the sleeves “almost never lose on the same days” and had near-zero correlations. That was a measurement artifact — the sleeves' return series were misaligned in time, which made them look independent when they weren't. Measured correctly, the three trend sleeves (VANGUARD, ORION, HELIOS) correlate 0.5–0.7 with each other: they are one trend bet expressed three ways, and bad trend months hit all three at once. The genuine diversifiers are BONDTREND, REVERSAL, TURN-OF-MONTH, and CRYPTO, whose correlations to the rest run from slightly negative to about 0.4. Diversification is real and it does smooth the ride — but it is modest, not magic.
v4 has exactly one automatic overlay:
Earlier versions of this page advertised two more brakes, and honesty requires explaining where they went. The 15% vol target was removed because, under the no-margin 100% cap, it can only ever reduce exposure — it was pure return drag with no crash protection the gate didn't already provide. The drawdown throttle was removed because a sign bug made the advertised formula inert (it almost never engaged), and the corrected version made results worse out-of-sample. The strategy never uses borrowed money or leverage above 100% — only the leverage built into the LETFs themselves.
A leveraged ETF magnifies the daily return of an index. TQQQ aims for 3× the daily return of NASDAQ-100. More bang per buck when right; bigger losses when wrong — which is exactly why we need good timing signals and safety brakes. Universe: U.S. stocks (UPRO, TQQQ, SSO, QLD), sectors (SOXL, TECL, FAS), real estate (DRN), international (EDC, YINN), commodities (UCO, UGL, NUGT), long bonds (TMF, UBT, TYD). Plus IBIT (Bitcoin) and BIL (cash).
A signal is generated every evening (after close, Mon–Fri). On most days there are 0–3 small trades. Annual turnover ~15×. Active but not high-frequency — one person can execute 2–3 market-on-open orders by hand each morning.
No strategy is guaranteed. PHOENIX can lose money. In the backtest (2014–2026), the worst peak-to-trough drop was about −17.9% — better than the S&P 500's historical −30 to −50% bear-market losses, but still meaningful, and a live drawdown can always be worse than the backtest's. An earlier claim that a synthetic 2008 stress test lost only −1.3% has been retired along with the flawed synthetic data behind it. Zero borrowed money — gross exposure is capped at 100% at every layer, with de-risked money parked in T-bills.
After U.S. close, an automated process pulls the latest prices for the ETFs and macro data (VIX, credit spreads, rates). It runs each of the seven strategies, blends them at equal weight, applies the tail gate, and produces the trade list above. You place those trades at tomorrow's open. That's it.
The technical version: weights, signals, the tail gate, execution. Nothing magical — everything mechanical, deterministic, and look-ahead-free.
PHOENIX blends 7 rule-based LETF trading strategies — each using different signals — at static equal weight. Seven sub-strategies — VANGUARD, ORION, HELIOS, CRYPTO, REVERSAL, TURN-OF-MONTH, BONDTREND — each get 1/7 of the book (1/6 before 2024-01-11, when CRYPTO activates at IBIT's listing). Nothing is fitted: there are zero estimated allocation parameters. The single overlay is a vol-regime tail gate that halves exposure when 60-day realized vol breaches its trailing 252-day 99th percentile (computed with a 2-day lag). Gross exposure is capped at 100% at every layer — no margin, no borrowing.
Full-window (2014–2026, net of costs): Sharpe 1.08, CAGR 22.2%, MDD −17.9%, Calmar 1.24, Sortino 1.39. The 2019–2026 window: Sharpe 1.09, CAGR 24.8% (SPY over the same window: 0.86 / 13.8% / −32% MDD). $10,000 invested at the start of 2014 becomes about $121,000 by mid-2026 (12.1×, no margin).
The seven sleeves (equal-weighted at 14.3% each, structurally different signals):
An earlier version of this page claimed the sleeves were “essentially uncorrelated (max pairwise |ρ| = 0.19)” and invoked a √N Sharpe uplift. Both claims were artifacts of misaligned return dating and are retracted. Measured correctly, the three trend sleeves — VANGUARD, ORION, HELIOS — correlate 0.5–0.7 with each other: one trend bet expressed three ways. The genuine diversifiers are BONDTREND, REVERSAL, TURN-OF-MONTH, and CRYPTO (pairwise correlations from −0.19 to roughly 0.4). The blend's Sharpe (~1.08) sits above the typical single sleeve, so diversification earns its keep — but modestly, not by a √N miracle.
The vol regime gate cuts exposure to 50% when 60-day vol breaches its trailing 252-day 99th percentile, computed with a 2-day shift — no look-ahead. That is the only overlay in v4. The 15% vol target was removed: under the hard 100% gross cap it could only subtract exposure and acted as pure drag. The drawdown throttle was removed: a sign bug made the shipped formula inert, and the corrected version hurt out-of-sample. Both removals are documented in alt/PHOENIX_V4_REVIEW.md.
Survivorship bias, transaction-cost sensitivity, weight perturbation (2,000 MC), the 1/N-vs-allocator comparison, block-bootstrap null (10,000 iter, p < 0.0001 full-sample), the QUANTUM retirement ablation, and a synthetic-extension test that was itself retired. Full-sample Sharpe 90% CI: [0.68, 1.50]; OOS lower bound 0.58. Honest live expectation: below the ~1.1 backtest — structure-level selection bias (this is the configuration that survived three audits) is disclosed, not modeled.
| Sub-strategy signals | Each of VAN/ORI/HEL/CRY/REV/TOM/BND computes its own target LETF weights at close T−1. VAN: monthly 189d-momentum + 200dma rotation over QLD/UGL/TMF/TYD with a macro trigger ladder. ORI: weekly (Wed) top-4 of 12 risk LETFs by 252d momentum + safe book. HEL: weekly (Fri) 42/189d underlying trend via LETFs. CRY: weekly (Fri) 63d momentum on BTC/ETH via IBIT/ETHA (from 2024-01-11). REV: daily z-score dip-buy signal, ~5-day holds. TOM: calendar — TQQQ over the last 4 + first 3 sessions. BND: weekly (Wed) TLT 63d momentum via TMF. |
| Blend weights | Static equal weight: 1/7 per active sleeve (1/6 before CRYPTO activates 2024-01-11). Zero estimated parameters — nothing is fit to returns. |
| Overlay · vol regime gate | The only overlay. Binary: 60-day vol > trailing 252-day 99th percentile (2-day shift) → mult = 0.5×; else 1.0×. Never leveraged above 100%. |
| Retired overlays | The 15% vol target and the −10% drawdown throttle from earlier versions are gone: the vol target was pure drag under the no-margin cap, and the throttle was inert due to a sign bug (its corrected form failed out-of-sample). |
| Execution | All signals/weights/multipliers at close T−1. Execute next-day open. 5 bps/side baseline TC + 10 bps per unit of daily leverage change (sensitivity to 35 bps/side in Test 02). |
| Daily cron | GitHub Actions, 6:30pm ET (23:30 UTC) Mon–Fri. Pulls yfinance + FRED, runs alt/live_signal.py, commits. |
Headline metrics, equity curve, drawdown, rolling Sharpe, and the in-sample vs out-of-sample test that mattered most.
| Window | Sharpe | CAGR | Ann Vol | MDD | Calmar | Sortino | NAV× |
|---|
| Strategy | Weight | Sharpe | CAGR | MDD |
|---|
| VAN | ORI | HEL | CRY | REV | TOM | BND |
|---|
When the strategy worked, when the brakes bit, and which sleeve carried which year.
The overlay is designed to bite rarely but firmly. The seven-sleeve blend delivers MDD −17.9% (2014+) with gross exposure ≤ 100% at all times; most day-to-day de-risking happens inside the sleeves themselves (safe books, cash positions). The only portfolio-level overlay is the 99th-percentile vol-regime gate, which halves exposure in extreme-vol regimes (Q1 2020 COVID, Q3 2022 dual stock-bond selloff). The old 15% vol target and drawdown throttle were removed in the v4 audit — one was pure drag, the other never actually worked.
| Sleeve | Weight | Cadence | Role |
|---|---|---|---|
| VANGUARD | 14.3% | Monthly | Trend (LETF rotation) |
| ORION | 14.3% | Weekly (Wed) | Trend (cross-sectional) |
| HELIOS | 14.3% | Weekly (Fri) | Trend (underlying) |
| CRYPTO (IBIT/ETHA) | 14.3% | Weekly (Fri) | Diversifier · active 2024+ |
| REVERSAL | 14.3% | Daily signal · ~5d hold | Diversifier (dip-buy) |
| TURN-OF-MONTH | 14.3% | Calendar (last 4 + first 3) | Diversifier · flat since 2024 |
| BONDTREND (TMF) | 14.3% | Weekly (Wed) | Diversifier · negative since 2019 |
| Sleeve | 30d ret | Contribution |
|---|
| Rank | LETF | 63d Total Return | Status |
|---|
| Year | VAN | ORI | HEL | CRY | REV | TOM | BND | TOTAL |
|---|
Survivorship, transaction-cost sensitivity, weight perturbation, the 1/N-vs-allocator comparison, block-bootstrap null, the QUANTUM retirement ablation, and one test that itself got retired. Tap each test to expand.
Added 15 LETFs that went to ~zero NAV (SOXS, SQQQ, SPXU, TECS, FAZ, YANG, EDZ, ERY, DRV, DUST, LABD, UVXY, VIXY, GLL, UNG) alongside the 17 survivors, re-ran the simplified momentum strategy. All 15 added names ended at 0.000–0.010× NAV — the literal definition of survivorship hit.
| Window | Survivors only | + Collapsed | Delta |
|---|
— momentum did pick collapsed names during their rally-then-fade phases. Real effect, quantified. A 0.10–0.20 Sharpe haircut for survivorship is an honest deduction.
PHOENIX embeds 5 bps/side TC at baseline (plus a leverage-change cost). Sweep additional TC to quantify slippage risk.
| Total TC (bps/side) | Ann drag | Full Sharpe | Full CAGR | OOS Sharpe |
|---|
At 25 bps/side (thin-LETF realistic), Sharpe 0.78 full / 0.84 OOS — still positive, no longer impressive. At 15 bps/side, 0.92 / 0.97. Each extra 5 bps/side costs roughly 0.07–0.08 of Sharpe.
±30% random multiplicative perturbation of all sleeve weights, renormalized (raw blend, before overlay). Tests whether results depend on the exact 1/7 allocation.
| Percentile | Full Sharpe | IS Sharpe | OOS Sharpe |
|---|
5th–95th percentile full-sample Sharpe: 1.03–1.07. The blend barely notices ±30% weight noise — as expected with equal weights and nothing fitted. The retired v3 walk-forward allocator scores higher on this in-era view, but lost to plain 1/N on every out-of-era window (see Test 04) — which is why it was retired.
Compares the shipped v4 equal-weight blend (zero estimated parameters) against the retired v3 walk-forward allocator (trailing-Sharpe budgets, annually refit) on the same sleeves and overlay.
| Scheme | Full Sharpe | IS Sharpe | OOS Sharpe |
|---|
The allocator looks better on the full in-era window; plain 1/N beat it on every out-of-era test window. Good in-era, worse out-of-era is the signature of a fitted allocator — which is why v4 ships static equal weight instead.
Circular block bootstrap with 21-day blocks. Tests (A) confidence interval around the point-estimate Sharpe and (B) p-value under the null that the strategy has zero expected return.
| Window | Observed Sharpe | 90% CI (5th–95th) | Std. Err | p(SR ≥ obs | null) |
|---|
Full-sample p < 0.0001; OOS p = 0.0001; IS p = 0.0038. The edge is statistically real — but the intervals are wide: full-sample CI [0.68, 1.50], OOS lower bound 0.58. That is compatible with a good-but-ordinary strategy, not a 2+ Sharpe one.
The original QUANTUM sleeve's published 2010–2018 history was in-sample XGBoost output — the model was evaluated on data it had memorized (full SR 1.73, IS 2.73, OOS 0.87). Rebuilt honestly with walk-forward training, it delivers standalone Sharpe ~0.2–0.6 with a −77% max drawdown; adding it back as an 8th equal-weight sleeve lowers the blend. It was retired in v3 and stays retired in v4.
| QUANTUM series | Standalone Full SR | Standalone OOS SR | Blend Full SR | Blend OOS SR |
|---|
Blend columns: the first row pairs the debunked published series with the shipped v4 blend that excludes QUANTUM; the second row shows the honest walk-forward rebuild and the blend with it re-added as an 8th sleeve — which drags full Sharpe from ~1.08 to ~0.98. Retirement was the right call.
Earlier versions of this page showed a 2005-extension built on synthetic LETF histories — including a claim that the strategy sailed through the 2008 GFC with a −1.3% drawdown. That test has been retired and its numbers are no longer quoted anywhere on this page.
| Backtest full Sharpe (2014–2026, net) | 1.08 |
| Backtest CAGR | 22.2% |
| Backtest max drawdown | −17.9% |
| Bootstrap p-value vs null (10k iter, full) | < 0.0001 |
| Bootstrap 90% CI (full / OOS lower bound) | [0.68, 1.50] / 0.58 |
| Weight perturbation ±30% | flat (±0.03) |
| v3 walk-forward allocator | retired — 1/N won out-of-era |
| QUANTUM ML sleeve | retired — in-sample output |
| Synthetic 2005 / GFC extension | retired — flawed builder |
| Survivorship-bias adjustment | −0.10 to −0.20 |
| TC @ 15 bps/side (realistic thin-LETF) | −0.16 |
| Structure-level selection bias | disclosed, not modeled |
| Honest live Sharpe expectation | below the ~1.1 backtest |
The suite supports “statistically real, moderately robust” — not “exceptional.” Bootstrap rejects pure luck (p<0.0001 full-sample), realistic costs and weight noise don't kill it, and everything that failed audit is retired rather than re-argued. The honest expectation for live performance is below the backtest: this is the configuration that survived three audits, and that survivorship is itself a selection effect we can disclose but not price. Zero portfolio margin or borrowing anywhere in the backtest or the live signal. Live record starts 2026-07-02.