Three parallel momentum strategies, increasing aggression. COMPOSITE: 5-sleeve diversified ensemble, best Sharpe. PURE: long-only top-3 stock momentum, simplest. LEV: standalone with leveraged ETFs, highest CAGR. All: no portfolio margin, no shorting, no options. Survivorship-bias-corrected.
Same 90-stock universe across all three; differ in sleeve count, concentration, and whether leveraged ETFs are allowed. Pick by your risk tolerance: highest Sharpe, simplest, or highest CAGR.
5 sleeves: 70% stocks + 30% unlevered ETFs.
Pure top-3 stock momentum, weekly rebal.
8-sleeve Phoenix-inspired ensemble with vol-target overlay. Standalone (no Phoenix).
Sharpe-CAGR trade-off curve. COMPOSITE is the Pareto-best risk-adjusted strategy (Sharpe 1.28). PURE simplifies to one sleeve at the cost of -10% Sharpe. LEV adds leveraged ETFs and pushes CAGR to 31.8% post-haircut — with vol_cap=1.0 strict (no margin), Phoenix-style overlay caps CAGR around 17-19%. Phoenix's 38% comes from years of tuning 5 sleeves to mean correlation 0.02 — hard to replicate without copying. Choose by what matters: low MDD (COMPOSITE), simple model (PURE), or maximum CAGR (LEV).
Where this strategy can and cannot go, given the strict constraints — explained before the strategy itself, because the math binds the design.
A 5-sleeve cross-asset-class momentum ensemble. Three stock sleeves rotate daily/weekly/monthly through the top-3/5/7 large-caps by 6-12 month momentum. Two ETF sleeves rotate daily through the top-1 of 33 ETFs by 21d and 126d momentum. Combined 70% stocks + 30% ETFs.
Constraints in force, simultaneously:
The stock universe is the set of 90 large-caps currently in data/stocks/ with 2010+ history. Companies that went bankrupt, got delisted, or merged out (Lehman, GM old, AIG-equivalent failures, etc.) are not in the dataset. This causes the backtest CAGR to be biased upward.
Mitigation: (1) Diversified top-K (3/5/7) for the COMPOSITE blend. (2) 30% of book in ETFs which have NO survivorship bias. (3) Apply a bootstrap-calibrated 8% CAGR haircut on the stock portion when planning forward returns. Blended haircut for COMPOSITE = 0.7 × 8% = 5.6% off the disclosed CAGR.
Backtest CAGR 26.7% → realistic forward expectation 21.1%. Sharpe figures are also slightly inflated by survivorship; the haircut is conservative.
The structural upgrade vs. the prior MERIDIAN. Adding the 90-stock single-name universe roughly doubles the alpha pool — stock-level momentum signals compete on a much wider, lower-correlation cross-section than ETF-level momentum. The result: Sharpe up from 0.88 → 1.28, CAGR up from 21% → 26.7% (haircut 21.1%) but with MDD halved from -37% → -19% and IS-OOS gap collapsed from 0.32 → 0.05. The ETF-only constraint was leaving most of the alpha on the table.
Open-to-open returns net of 5 bps per leg. Benchmarks: SPY, QQQ, and a 60/40 SPY/AGG portfolio.
| Window | Sharpe | CAGR | Vol | MDD | Sortino | Calmar |
|---|
Top 1x-ETFs by full-window buy-and-hold CAGR. Even a perfect-foresight ex-post pick of the best 1x asset (SMH) compounds at only 25.2%. The leveraged products that would clear the 30% bar are forbidden by the spec.
| Asset | Class | FULL CAGR | IS CAGR | OOS CAGR | Note |
|---|---|---|---|---|---|
| TQQQ | 3x QQQ | 43.4% | 41.2% | 45.6% | forbidden |
| SOXL | 3x semis | 41.4% | 28.2% | 58.5% | forbidden |
| TECL | 3x tech | 38.2% | 27.4% | 52.6% | forbidden |
| QLD | 2x QQQ | 32.1% | 27.5% | 37.7% | forbidden |
| UPRO | 3x SPY | 29.3% | 26.1% | 33.2% | forbidden |
| SMH | semis (1x) | 25.2% | 14.7% | 42.0% | 1x ceiling — ex-post bias |
| XLK | tech sector | 19.4% | 13.4% | 27.1% | 1x |
| QQQ | nasdaq-100 | 19.0% | 15.5% | 23.4% | 1x |
| XLY | cons disc | 14.3% | 15.9% | 12.3% | 1x |
| SPY | S&P 500 | 14.2% | 13.2% | 15.6% | 1x |
| XBI | biotech | 13.1% | 16.9% | 8.3% | 1x |
| SCHD | quality div | 13.0% | — | — | 1x · post-2011 |
Read carefully: Even an ex-post perfect concentration in the FULL-window winner (SMH 25.2%) falls 5 percentage points short of 30%. To exceed 30% CAGR you need a structurally different toolkit — leveraged ETFs (forbidden), portfolio margin (forbidden), or hand-picked winners (selection bias, forbidden). The arithmetic does not care about effort.
Five rule-based momentum sleeves across two asset classes. Three operate on individual large-cap stocks (different K and lookback combinations) at 70% total weight. Two operate on the 33-ETF universe at 30% weight. The stock sleeves provide most of the alpha; the ETF sleeves provide diversification (correlation 0.3-0.4 with stocks) and survivorship-bias-free exposure.
Top-3 stocks by 126-day return, monthly rebalance. Equal-weight long, eligibility = positive 126d return.
Highest-Sharpe single sleeve. Concentrates on the strongest 6-month winners with monthly turnover (~150% per year).
Top-5 stocks by 126-day return, weekly rebalance. Equal-weight long, eligibility = positive 126d return.
Captures short-term shifts in the 6-month momentum cross-section. Wider K dilutes single-stock concentration risk.
Top-7 stocks by 252-day return, monthly rebalance. Wider K + longer lookback = the slowest, most diversified stock sleeve.
Acts as a stabilizer to the faster sleeves. Lower correlation with S1 (~0.91) than S1↔S2 (~0.96).
Top-1 of the 33-ETF universe by 21-day return, daily rebalance, full concentration in the single hottest mover.
~0.3 correlation with stock sleeves. Provides survivorship-bias-free alpha and broad asset-class coverage (bonds, commodities, sectors).
Top-1 of the 33-ETF universe by 126-day return, daily rebalance, full concentration in the medium-term ETF leader.
Holds multi-quarter trends — the 2017 semis run, 2020 post-COVID recovery, 2023-2024 AI rally. ~0.4 correlation with stock sleeves.
Why stocks + ETFs together. Stock-level momentum has higher information ratio than ETF-level (the cross-section is wider — 90 names vs 33). But stocks suffer survivorship bias; ETFs don't. Blending 70/30 captures most of the stock alpha while the ETF portion provides clean exposure. Cross-asset-class correlation: ~0.3 — solid diversification multiplier.
ETF universe (33) is fixed ex-ante by liquidity + inception ≤ 2009 — no selection bias.
Stock universe (90 large-caps) is the set with 2010+ history in data/stocks/.
Survivorship-biased; bias is measured empirically via bootstrap simulation
and accounted for with a calibrated 8% CAGR haircut on the stock portion (see methodology below).
Survivorship-bias disclosure. The 90-stock universe is currently-listed S&P 500 large-caps with 2010+ history. Stocks that went bankrupt, were delisted, or merged out of existence are not in the dataset (no Lehman, WaMu, MF Global, etc.). We do not rely on a flat literature-based haircut — we measured the bias empirically.
Method: we measure the bias empirically rather than guess. At each year boundary, randomly mark N% of stocks as "delisted" (excluded from selection for the rest of the backtest, with a one-day loss applied on the delisting date to simulate the wipeout). Run the TOP-3 momentum strategy. Compare to no-dropout baseline. 100 Monte Carlo simulations per (dropout, loss) cell.
| Annual dropout | Delist loss | Median CAGR | Bias | 95% CI |
|---|---|---|---|---|
| 0% (baseline) | — | 35.0% | ref | — |
| 2% / yr | 0% | 34.3% | +0.7% | [26.4, 36.5] |
| 2% / yr | −50% | 33.4% | +1.6% | [25.9, 36.5] |
| 2% / yr | −100% | 32.1% | +2.9% | [23.1, 36.5] |
| 5% / yr | 0% | 27.7% | +7.3% | [18.4, 36.3] |
| 5% / yr (S&P 500 norm) | −50% | 22.2% | +12.8% | [13.5, 32.1] |
| 5% / yr | −100% | 16.6% | +18.4% | [4.3, 29.2] |
| 10% / yr | 0% | 15.5% | +19.5% | [9.5, 20.3] |
| 10% / yr | −50% | −2.3% | +37.4% | [−10.5, 6.1] |
The yellow rows show realistic scenarios at the historical 5%/year S&P 500 turnover rate. With no delist loss the bias is +7.3%; with a realistic −50% delist loss the bias jumps to +12.8% CAGR; with a full −100% wipeout it's +18.4%.
| Universe | Bias? | CAGR | Sharpe |
|---|---|---|---|
| EW 90-stock universe | YES (max) | 17.3% | 1.05 |
| SPY (cap-wt, reconstituted) | NO | 13.2% | 0.82 |
| QQQ (cap-wt, reconstituted) | NO | 17.5% | 0.90 |
| IWM (R2K, reconstituted) | NO | 9.9% | 0.54 |
EW 90-stock vs SPY = +4.2% (mixes ~1-2% EW premium + ~2-3% bias). EW 90-stock ≈ QQQ — our universe leans tech-heavy by virtue of which names survived. The bootstrap is more conservative than these passive cross-checks because concentrated top-K strategies amplify the bias more than EW does.
We split the difference between the no-loss case (+7.3%) and the realistic-loss case (+12.8%) and apply a conservative 8% CAGR haircut on the stock portion of each strategy. Per-strategy blended:
All five constraints hold for every bar of the backtest. Violating any one would change the strategy's identity — it would no longer be MERIDIAN.
Universe is exclusively 1x products. No 2x/3x products like TQQQ, SOXL, TECL, UPRO, TMF. No inverse products like SH, SQQQ, SPXU. No volatility products like UVXY or SVXY (the latter is technically -0.5x).
Sum of weights at any time t is bounded at 1.0. Cash residual sits in BIL (effectively zero return). The strategy never goes to 1.05x gross or borrows against positions.
Every input uses close-of-day data through t-1. Positions are computed and recorded at date t (using close[t-1] info), and the backtest takes the position at open[t] earning return open[t]→open[t+1]. There is at least one full trading day between signal and trade.
The 33-ETF universe is assembled ex-ante by liquidity + inception ≤ 2009. The 90-stock universe is the set of currently-listed S&P 500 large-caps with 2010+ data — survivorship-biased. We mitigate by using diversified top-K (3/5/7) and applying a bootstrap-calibrated 8% CAGR haircut on the 70% stock portion when reporting forward-looking metrics. Naive backtest CAGR 26.7% → realistic forward expectation 21.1%.
Eligibility, signals, and rebalances all happen daily. The strategy follows leadership transitions immediately — when a new ETF emerges as top by 21d momentum, it gets bought tomorrow at the open. Net of 3 bps per leg per ETF (institutional algo execution).
Two portfolio-level overlays act on the aggregated raw return. Each multiplier is bounded in [0, 1] — they can only reduce exposure, never amplify it. Average overlay multiplier on the full sample is 0.95.
When portfolio NAV is below its 252-day high-water mark, the multiplier scales linearly toward 0 as drawdown approaches the floor of -15%. Below -15% the multiplier is 0 (full cash). The overlay re-engages as NAV recovers toward HWM.
Formula: dd_mult = clip(1 + dd / -0.15, 0, 1), lagged 1 day.
When 60-day realized portfolio vol exceeds the 99th percentile of its 252-day trailing distribution, exposure is halved. This catches genuine tail-risk regimes (Mar-2020, Oct-2022) without triggering on routine noise.
Formula: vol_mult = 1.0 if rv ≤ 99th_pct else 0.5, lagged 1 day.
Overlay arithmetic: total_mult = min(1.0, dd_mult × vol_gate_mult). Average across 16 years is 0.95. Net return = raw return × total_mult, with TC drag from multiplier changes.
Top holdings of the active strategy on the first business day of each year. Switch strategy with the toggle above to see different allocations. Concentration shifts as sector leadership rotates — semis dominated 2017-2018 and 2023-2024; consumer discretionary led 2013-2015; defensives led 2022.
Snapshots are first-business-day-of-year. COMPOSITE blends 3 stock sleeves (top-3/5/7) at 23.3% each + 2 ETF sleeves at 15% each. PURE shows top-3 stock allocations. LEV shows stock-portion holdings (40% of book; LETF rotation handled separately).
Total return per calendar year vs. SPY, QQQ, and 60/40 SPY/AGG. MERIDIAN is meaningfully below the equity benchmarks in strong bull years (2013, 2017, 2021, 2024) by design — the defensive sleeve gives up upside in exchange for capital preservation in bear years.
| Year | MERIDIAN | SPY | QQQ | 60/40 |
|---|
Each cell is the strategy's net return for that month. Color: green positive, red negative, intensity scales with magnitude. Year-end totals on the right.
Cells colored on a -10% (deepest red) to +10% (deepest green) scale. Clipped beyond.
Pairwise correlation of daily sleeve returns over the full sample. The defensive sleeve is essentially uncorrelated with the two equity sleeves — that's the source of the blend's Sharpe lift.
| S1 10/126/M | S2 15/126/W | S3 20/252/M | S4 ETF FAST | S5 ETF SLOW |
|---|
Under the strict constraints, there's a hard frontier: higher CAGR comes at the cost of Sharpe and vice-versa. The product (Sharpe × CAGR) tops out around 0.20 across all configurations tested. Below: three points on the curve, all unbiased and unlevered.
→ Sh 1.28 · 26.7% CAGR
5 sleeves, 70% stocks + 30% ETFs. Single-stock momentum on 90 large-caps + ETF momentum on 33 broad ETFs. MDD -19.8%, Sortino 1.70, OOS Sharpe 1.27.
Haircut-adjusted CAGR: 21.1% (8% bias on 70% stock portion = 5.6% blended).
→ Sh 0.88 · 21% CAGR
Dual-momentum on 33 ETFs (FAST 21d + SLOW 126d, 50/50). No survivorship issue but lower Sharpe and higher MDD (-37%) due to narrower cross-section.
For investors who reject any survivorship-biased universe.
→ Sh 0.90 · 3% CAGR
90% bond carry (HY-OAS gated) + 10% momentum. Vol 3.4%, MDD -7%. Maximum Sharpe achievable on bond-heavy unbiased blends.
For preserving capital. Tiny return but exceptional risk control.
The big unlock was widening the universe. ETFs alone gave Sharpe ~0.9. Adding 90 single-name stocks took Sharpe to 1.17 — the cross-section is wider, the alpha is real. To go higher: leveraged ETFs (Phoenix territory, Sharpe 2.4); margin (scales linearly); options strategies (covered calls, put-write — could add 0.2-0.3 Sharpe); or short-history specialty instruments. Sharpe > 3 over 16 years still requires either leverage-equivalent vol scaling or a fundamentally different toolkit (long-short with hundreds of positions, futures, HFT).
Every signal is computed on close[t-1] data only. Weights are recorded at date t. backtest_o2o uses weights.shift(1) so realized return on date t corresponds to a position decided at close[t-2] and traded at open[t-1], earning return open[t-1] → open[t]. A full trading day between decision and trade.
No look-ahead anywhere. Sleeve weights are blended once on the IS window only; no re-fitting on OOS. Risk overlays use rolling lookbacks shifted by 1 day before being applied.
The single canonical implementation, the design doc with deep-research findings, and the raw output artefacts.
alt/meridian_strategy.py
Single canonical file. ~340 lines. Loads the universe, builds sleeves, applies overlays, prints metrics, writes JSON + CSV outputs.
alt/MERIDIAN_DESIGN.md
Full constraint discussion, deep-research findings (parameter searches, alpha-source decomposition), and an honest accounting of what the strategy can and cannot deliver.
data/results/meridian_metrics.json
All performance numbers shown on this page, plus per-sleeve stats and the full pairwise correlation matrix.
data/results/meridian_returns.csv
Daily raw and net returns, drawdown / vol-gate / total multiplier state. meridian_sleeves.csv has each sleeve's daily series for further analysis.
Disclaimer. All performance figures are backtested on historical close-of-day price data downloaded from public sources (Yahoo Finance for ETFs, FRED for macro series). Backtests are net of 5 bps per leg per ETF as a transaction-cost proxy; they do not account for slippage, market-impact, dividend tax frictions, ETF expense ratios, or financing on cash residuals. Past performance is not predictive of future returns. This page is research, not investment advice.