1. Executive Verdict: Did the 4-Stage Funnel Beat the Market on Skill or Luck?
Across the 34-quarter out-of-sample walk-forward simulation (2017–2025) with Marcos López de Prado target-label purging (63d embargo) and 15 bps round-trip transaction friction across the 137-asset multi-sector & 3-tier historic IPO universe, the 🏰 7 Powers Screen + Overreaction Kelly (4-Stage Funnel) delivered +324.72% cumulative return (+18.55% CAGR) versus +154.99% (+11.66% CAGR) for the SPY S&P 500 benchmark.
Key Quantitative Takeaway: Approximately 75% of the outperformance (+522 bps/yr over SPY) comes from Stage 0 & Stage 1 Left-Tail Elimination (filtering out unseasoned IPO hype and structural moat erosion), while 25% (+167 bps/yr) comes from Stage 2 & Stage 3 Overreaction Quarter-Kelly Sizing. The strategy wins not by chasing high-beta moonshots, but by refusing to own decaying franchises during tightening regimes (-0.68% in 2022 vs. -14.09% SPY and -31.04% Rejected Moat Traps).
| Strategy Cohort | Total Return | CAGR | Sharpe | Sortino | Max Drawdown | Win Rate vs SPY | Jensen’s Alpha |
|---|---|---|---|---|---|---|---|
| 🏰 7 Powers + Overreaction Kelly (Stage 0–3) | +324.72% | +18.55% | 0.85 | 1.40 | -22.72% | 61.8% | +6.72%/yr |
| 🛡️ 7 Powers + 3M T-Bill Barbell | +265.49% | +16.47% | 0.87 | 1.43 | -18.93% | 58.8% | +6.03%/yr |
| 🏰 Stage 1 Equal-Weight (Moat Screen Only) | +276.38% | +16.88% | 0.86 | 1.38 | -15.78% | 58.8% | +5.22%/yr |
| 🤖 Quant ML Top Quintile (Q5 Unscreened) | +225.79% | +14.93% | 0.58 | 1.24 | -40.87% | 55.9% | +3.27%/yr |
| 📈 S&P 500 Benchmark (SPY) | +154.99% | +11.66% | 0.53 | 0.78 | -21.11% | — | 0.00% |
| ⚠️ Rejected Fallen Angels & Broken IPOs | +131.03% | +10.35% | 0.25 | 0.52 | -51.61% | 44.1% | -8.68%/yr |
| 📉 Quant ML Bottom Quintile (Q1) | -7.08% | -0.86% | -0.19 | -0.26 | -61.81% | 32.4% | -12.52%/yr |
2. Attribution Decomposition: What Genuinely Worked vs. Noise
A. Stage 0 & Stage 1 Moat Screening (+522 bps/yr Structural Alpha)
- Funnel Compression: At an average quarterly rebalance date,
101active tickers enter Stage 0,90survive the IPO lock-up seasoning filter (>= 378calendar days post-IPO and drawdown> -40%from post-IPO ATH),31pass the point-in-time cross-sectional 7 Powers threshold (M >= 62with non-collapsing126dmoat trajectory>= -8), and12receive active Quarter-Kelly capital allocation. - Fallen-Angel Ejection (
-8.68%/yrJensen’s Alpha Avoided): Dynamically ejecting decaying incumbents and broken growth stories—such asINTC(M=26.9),BA(M=28.9),DIS(M=26.9),BABA(M=27.9),DOCU(M=26.8),KSS(M=25.9),SPCE(M=26.8), andBYND(M=28.8)—compressed Maximum Drawdown from -40.87% (unscreened ML Q5) down to -15.78% (Stage 1 Equal-Weight) and -18.93% (Barbell).
B. Stage 2 & Stage 3 Overreaction Kelly Sizing (+167 bps/yr Incremental Alpha)
- Comparing 7 Powers + Overreaction Kelly (
+18.55%CAGR,-22.72%MDD) against Stage 1 Equal-Weight (+16.88%CAGR,-15.78%MDD) reveals the exact trade-off of active conviction sizing: concentrating up to12%per position into oversold wide-moat pullbacks adds +167 bps/yr of annualized return (+48.3% cumulative wealth over 8.5 years) at the cost of +6.94% deeper maximum drawdown. - Meanwhile, the 3M T-Bill Hurdle Barbell (
+16.47%CAGR,0.87Sharpe,1.43Sortino) achieves the highest risk-adjusted efficiency in the entire study by parking unallocated capital in risk-free Treasury bills (macro_yield_3m) whenever high-moat equities become overbought (RSI_14 >= 75).
C. 34-Quarter Factor Attribution by Dominant Helmer Power
Decomposing every Stage 1 survivor by its highest point-in-time normalized pillar confirms which structural moats actually compound:
| Dominant Helmer Power Pillar | Avg Quarterly Picks | Annualized CAGR | Avg Quarterly Return | Hit Rate (>0%) | Max Drawdown |
|---|---|---|---|---|---|
⚙️ Process Power & Scale Economies (NVDA, LLY, TSM, CAT, COST) |
8.3 | +20.00% | +5.33% | 67.6% | -22.27% |
🔒 Switching Costs (MSFT, ORCL, NOW, CRM, ISRG) |
6.8 | +12.91% | +3.77% | 67.6% | -15.88% |
💎 Cornered Resource & Branding (AAPL, UNH, REGN, NVO) |
6.4 | +10.01% | +2.74% | 70.6% | -12.20% |
🌐 Network Economies & Counter-Positioning (META, GOOGL, NFLX, UBER) |
9.5 | +9.33% | +3.18% | 61.8% | -34.10% |
3. Regime Diagnosis: Why the Model Trailed in 2018, 2019, and 2021
| Year | 🏰 7 Powers Kelly | 🛡️ 7 Powers Barbell | 🏰 Stage 1 Equal-Wt | 🤖 Quant ML Q5 | 📈 SPY Benchmark | ⚠️ Rejected Traps |
|---|---|---|---|---|---|---|
| 2017 (3Q) | +19.33% | +13.73% | +13.76% | +12.71% | +9.97% | +8.47% |
| 2018 | -11.33% | -9.71% | -9.00% | -15.30% | -5.82% | -12.66% |
| 2019 | +19.16% | +18.43% | +22.41% | +19.83% | +27.79% | +19.65% |
| 2020 | +47.20% | +41.15% | +26.40% | +69.64% | +16.17% | +54.78% |
| 2021 | +5.47% | +7.26% | +23.81% | +11.67% | +18.51% | +16.13% |
| 2022 | -0.68% | -0.28% | -5.81% | -22.61% | -14.09% | -31.04% |
| 2023 | +24.80% | +23.22% | +25.72% | +34.50% | +21.11% | +35.09% |
| 2024 | +39.04% | +35.42% | +24.47% | +35.58% | +20.06% | +19.01% |
| 2025 (3Q) | +27.84% | +23.09% | +37.98% | +23.62% | +13.38% | +29.38% |
- The
2021Euphoria Penalty (+5.47%Kelly vs.+23.81%Stage 1 Equal-Weight): In low-volatility stimulus melt-ups (2021), Stage 1 selected the right compounders (+23.81%), but Stage 2’s0.35xoverbought penalty (RSI_14 >= 75) prematurely cut winners while Stage 3 reallocated into value laggards experiencing temporary pullbacks (RSI_14 < 45). - The
2022Tightening Triumph (-0.68%Kelly vs.-14.09%SPY &-31.04%Traps): When the Federal Reserve hiked rates aggressively in2022, Stage 0/1 ejected unprofitable tech and broken IPOs while Stage 3 demanded a rising 3M T-Bill hurdle (macro_yield_3m), preserving nearly 100% of capital.
4. The Hard Audit: 5 Methodological & Structural Vulnerabilities
Despite eliminating manual ticker selection and De Prado target leakage, five structural flaws remain in the v1 implementation:
- Price/Volume Proxies Masquerading as Accounting Moats: Because historical point-in-time SEC XBRL statements (
Gross Margin,ROIC,R&D/Rev,Deferred Revenue) are not stored in the 12-year daily parquet panel, the 7 Powers scores are approximated from price/volume behavior (ret_252d,dist_sma_200,vol_63d,cmf_20,obv_slope_20). During commodity spikes (e.g.,NEMin 2020 orOXY/SLBin 2022), cyclical price momentum can temporarily mimicProcess PowerandScale Economies. - Static
is_whale_holdingBroadcast Leakage: Insrc/historical_dataset.py,is_whale_holdingandwhale_manager_countare populated fromdata/whales.yamland broadcast across a ticker’s historical rows (weight 0.25in theCornered Resourcepillar). While modest (~3.75%of totalM_i,t), historical backtests should either use point-in-time SEC 13F-HR filing dates or zero out static whale flags. - Quarterly (
63-Day) Rebalance vs.14-DayOverreaction Horizon Mismatch:RSI_14 < 45andbb_pct_b < 0.25capture 5-to-15 trading day liquidity shocks, whereas the walk-forward engine only rebalances once every 63 trading days. Mean-reversion velocity decays substantially after day 25. - Unenforced Intra-Quarter
2x ATRStop-Loss: Stage 3 sizes positions using an asymmetric payoff ratiob_i,tassuming downside is capped atmax(2 * ATR_14, 5%), yet the quarterly simulation evaluates rawtarget_return_3mat day 63 without triggering an intra-quarter exit whentarget_max_drawdown_3mbreaches-2 * ATR_14. - Delisted Bankruptcy Censorship in Yahoo Finance: While our 137-ticker universe includes surviving busts (
SPCE,BYND,PTON,LCID,RIVN,WBA,INTC), completely delisted liquidations (SIVB,FRC,BBBY,WE) are absent from Yahoo Finance’s active feed—meaning theRejected Fallen Angelscohort (+10.35%CAGR,-51.61%MDD) still slightly overstates the return of buying moat traps.
5. Phase 2 Engineering Roadmap
- Fix 1 (Zero Static Whale Leakage): Strip
is_whale_holdingandwhale_manager_countfrom historical backtest moat scoring insrc/ml_backtest.pyuntil point-in-time EDGAR 13F-HR timestamps are ingested. - Fix 2 (Enforce Intra-Quarter
2x ATRStop-Outs): Cap quarterly realized losses insrc/ml_backtest.pyat-2 * ATR_14 - 15 bpsslippage whenevertarget_max_drawdown_3mbreaches the stop level, aligning realized execution with the Quarter-Kelly denominator. - Fix 3 (
21-DayTactical Overreaction Rebalance Within63-DayMoat Refresh): Recompute the slow-moving Stage 1 7 Powers universe quarterly (63d), but execute Stage 2/3 overreaction entries and euphoria trims monthly (21d) to match the half-life of short-termRSI_14dislocations.