# QC Comparative Backtests — Strategy Catalog

**Issue** : See #1630
**Generated** : 2026-06-04 (automated catalog, pending live backtests)
**Updated** : 2026-06-11 — +3 baselines (MeanReversion v5.2 IBKR, AdaptiveAssetAllocation, PairsTrading). MeanReversion promoted Tier 2 → Tier 1 (0.29 → 0.81). AAA promoted Tier 4 → Tier 1 (untested → 0.509).
**Post-#2801 verification** : 2026-06-14 — campaign started. The remediation #2801 (brokerage
model on 52 strategies via Lot 1 #2812, real fees via Lot 2 #2823/#2864, fixed end-dates via
Lot 3 #2813, OOS splits via Lot 5 #2824) **invalidates pre-remediation Sharpes** when the
brokerage/fee model was previously the negligible default. Entries marked
`✓post-#2801` in the **Verified** column are re-run live via MCP qc-mcp; all others remain
pre-remediation catalog values and need re-backtest before comparative conclusions.
**Methodology** : Period common 2018-01-01 → 2024-12-31 (US equities/multi-asset), 2020-01-01 → 2024-12-31 (crypto). Metrics from `projects/catalog.json` + prior backtest runs. Pending: standardized re-backtest via MCP qc-mcp.

---

## Legend

| Column | Definition |
|--------|------------|
| **Sharpe** | Annualized Sharpe ratio (net of costs where available) |
| **CAGR** | Compound annual growth rate (%) |
| **MaxDD** | Maximum drawdown (%) — *pending standardized re-run* |
| **Type** | `IND` = indicator, `ML` = machine learning, `DL` = deep learning, `RL` = reinforcement learning, `RISK` = risk-parity/vol-targeting, `OPT` = options, `COMP` = composite/multi-alpha |
| **Class** | Asset class |
| **Status** | `robuste` = Sharpe > 0.5, backtested; `historique` = Sharpe 0-0.5; `exploratoire` = negative Sharpe; `untested` = no backtest yet |

---

## Tier 1 — Robuste (Sharpe > 0.5, 41 strategies)

Strategies with solid risk-adjusted returns. These are the primary candidates for the comparative backtest baseline.

| # | Project | Type | Class | Sharpe | CAGR% | MaxDD% | Calmar | Status |
|---|---------|------|-------|--------|-------|--------|--------|--------|
| 1 | LongShortHarvest-QC | ML | Equities | ~~3.39~~ → **1.64** ✓post-#2801 | 45.6 | 17.0 | 2.68 | robuste (confirmed, **PSR 98.7%** top — catalog 3.39 was 1Y-OOS, full decade -52%) |
| 2 | HighBookToMarketFScore-QC | ML | Equities | ~~2.09~~ → **0.41** ✓post-#2801 | 14.5 | 60.4 | 0.24 | **historique** (downgraded: real fees, MaxDD -60%) |
| 3 | PuppiesOfTheDow-QC | IND | Equities | ~~1.99~~ → **0.30** ✓post-#2801 | 9.6 | 28.8 | 0.33 | **historique** (downgraded: real fees) |
| 4 | LeveragedETFMomentum-QC | IND | Equities (lev ETF) | ~~1.80~~ → **1.78** ✓post-#2801 | 126.4 | 53.3 | 2.37 | robuste (confirmed, leveraged — MaxDD -53% expected) |
| 5 | Positive-Negative-Splits-ML | ML | Equities | ~~1.74~~ → **1.51** ✓post-#2801 | 75.7 | 37.6 | 2.01 | **non robuste** (revu par #19242 : reproduction 2018 → 2024-04 Sharpe 1,16 / PSR 43,6 %, 89 % du résultat en 2023-2024, capacité estimée par QC 3 000 dollars ; 2018 → 2026-09 : ruine en vente à découvert, variante long seul `NO BEATS`) |
| 6 | DynamicVIXSpyRegime-QC | ML | Equities/VIX | ~~1.72~~ → **1.00** ✓post-#2801 | 17.9 | 16.5 | 1.09 | robuste (confirmed, **PSR 69.4%** — catalog 1.72 was 1Y-OOS, full decade -42%) |
| 7 | MacroFactorRotation-QC | ML | Multi-asset | ~~1.23~~ → **0.73** ✓post-#2801 | 22.6 | 42.0 | 0.54 | robuste (revised Sharpe -40%, real fees) |
| 8 | Framework_Composite_TrendWeather | COMP | Equities | ~~1.16~~ → **1.14** ✓post-#2801 | 27.1 | 27.7 | 0.98 | robuste (confirmed, PSR 77.9%) |
| 9 | Trend-Following | IND | Equities | ~~1.07~~ → **0.41** ✓post-#2801 | 7.9 | 14.6 | 0.54 | **historique** (downgraded: real IBKR fees) |
| 10 | Multi-Layer-EMA | IND | Crypto (BTC) | ~~0.93~~ → **0.80** ✓post-#2801 | 25.0 | 57.1 | 0.44 | robuste (confirmed -14%, ML crypto holds, PSR 23.9%, MaxDD -57% BTC) |
| 11 | Portfolio-Optimization-ML | ML | Multi-asset | ~~0.90~~ → **0.88** ✓post-#2801 | 27.2 | 41.6 | 0.65 | robuste (confirmed -2%, monthly rebalance + fee-homogeneous US equity basket = near-immune, PSR 37.2%) |
| 12 | EMA-Cross-Stocks | IND | Equities | ~~0.87~~ → **0.99** ✓post-#2801 | 29.2 | 35.7 | 0.82 | robuste (PSR 49.7%, borderline significant) |
| 13 | CausalEventAlpha | ML | Equities | ~~0.78~~ → **0.45** ✓post-#2801 | 11.7 | 38.8 | 0.30 | **historique** (downgraded -43%, monthly cadence but full-basket rotation = high turnover, PSR 5.5%) |
| 14 | Gaussian-Direction-Classifier | ML | Equities | ~~0.76~~ → **0.76** ✓post-#2801 | 23.1 | 25.6 | 0.90 | robuste (confirmed 0%, daily schedule but low realized turnover on liquid mega-cap basket = near-immune, PSR 22.7%) |
| 15 | ML-Temporal-CNN | DL | Equities (QQQ) | ~~0.73~~ → **0.46** ✓post-#2801 | 12.5 | 30.8 | 0.41 | **historique** (downgraded -37%, DL/CNN overfits real fees, PSR 5.2%) |
| 16 | TrendStocksLite | IND | Equities | ~~0.72~~ → **0.71** ✓post-#2801 | 18.0 | 33.7 | 0.53 | robuste (confirmed -2%, weekly trend on 15 liquid large-caps = low realized turnover, near-immune, PSR 25.0%) |
| 17 | ML-LLM-Summarization | ML/NLP | Equities | 0.69 | 15.5 | — | — | robuste |
| 18 | ML-RandomForest | ML | Equities | ~~0.68~~ → **0.70** ✓post-#2801 | 20.6 | 40.9 | 0.50 | robuste (confirmed +3%, bi-weekly RF on 10 mega-caps = moderate turnover near-immune on fee-homogeneous US equity basket, PSR 18.4%) |
| 19 | AllWeather | RISK | Multi-asset | ~~0.67~~ → **0.47** ✓post-#2801 | 7.5 | 16.4 | 0.46 | **historique** (downgraded -30%, low-turnover multi-asset NOT near-immune, PSR 19.6%) |
| 20 | ML-Trend-Scanning | ML | Multi-window | ~~0.66~~ → **0.33** ✓post-#2801 | 7.1 | 29.4 | 0.24 | **historique** (downgraded -50%, daily rebalance on SPY/TLT/GLD multi-asset = very high turnover crushed by real IBKR fees, PSR 7.8%) |
| 21 | VolTarget-Momentum | COMP | Multi-asset | ~~0.65~~ → **0.50** ✓post-#2801 | 11.1 | 21.2 | 0.53 | robuste borderline (revised -23%, PSR 9.4%) |
| 22 | SectorMomentum | IND | Equities+Bonds+Gold | ~~0.62~~ → **0.56** ✓post-#2801 | 13.1 | 22.8 | 0.58 | robuste (mild -9%, monthly rebalance winner-takes-all on 3 ETFs = low rotation near-immune, contrasts heavier multi-sleeve baskets; PSR 13.8% low, not a true leader) |
| 23 | BlackLitterman-Momentum | COMP | Equities/ETF | ~~0.60~~ → **0.83** ✓post-#2801 | 15.8 | 16.9 | 0.94 | **robuste (catalog CORRECTED UPWARD +38%)** — catalog 0.60 was stale; code already had IBKR so the docstring v2 BEST (0.823, CAGR 15.7%, MaxDD -16.9%) was the WITH-fees value; monthly rebalance + fee-homogeneous 15-stock US equity basket = near-immune to fees; **PSR 51.4% = true leader (borderline significant)** |
| 24 | Crypto-MultiCanal | IND | Crypto (BTC) | ~~0.58~~ → **0.33** ✓post-#2801 | 4.6 | 14.1 | 0.33 | **historique** (downgraded -43%, crypto indicator NOT robust post-fees, PSR 13.0%) |
| 25 | ML-FeatureEngineering | ML | Equities (10 mega-caps) | ~~0.57~~ → **0.65** ✓post-#2801 | 18.7 | 28.1 | 0.67 | robuste (catalog CORRECTED UPWARD +15% — was stale; real Sharpe of current RF+GB 18-feature ensemble near-immune on fee-homogeneous US equity; PSR 15.1% low, not a true leader; universe corrected Multi-asset→Equities; baseline-clone 32952140) |
| 26 | Markov-Regime-Detection | ML | Multi-asset (SPY/TLT/GLD) | ~~0.57~~ → **0.375** ✓post-#2801 | 8.4 | 24.4 | 0.35 | **historique** (downgraded -34% — catalog 0.57 was STALE; IBKR fees only ~8% drag from intrinsic ~0.41; binary 80% SPY↔TLT swap per regime change = high per-event turnover despite low frequency; PSR 5.8%) |
| 27 | ML-XGBoost | ML | Equities (15 mega-caps) | ~~0.57~~ → **0.555** ✓post-#2801 | 14.5 | 40.4 | 0.36 | robuste (confirmed mild -3%, bi-weekly GradientBoostingRegressor on fee-homogeneous 15 mega-caps = near-immune; PSR 10.6% low, not a true leader; universe corrected Multi-asset→Equities) |
| 28 | MomentumStrategy | IND | Equities | ~~0.57~~ → **0.50** ✓post-#2801 | 11.2 | 25.8 | 0.43 | robuste borderline (at threshold -12%, PSR 9.3% non-significant) |
| 28b | MeanReversion | IND | Equities (sectors) | ~~0.81~~ → **0.81** ✓post-#2801 | 10.0 | 7.5 | 1.34 | robuste (confirmed, PSR 46.8% near-significant, low-turnover multi-asset holds = signal-frequency immunity) |
| 29 | RegimeSwitching | ML | Multi-asset (SPY/QQQ/IEF/GLD) | ~~0.55~~ → **0.581** ✓post-#2801 | 12.3 | 33.0 | 0.37 | robuste (confirmed mild +6%, multi-asset HOLDS via explicit turnover suppression — regime-change trigger + anti-micro-rebalancing + beta-annealing; contrasts AllWeather -30% multi-sleeve; PSR 7.0% low not a true leader) |
| 30 | Temporal-CNN-Prediction | DL | Equities (QQQ top-3) | ~~0.54~~ → **0.161** ✓post-#2801 | 4.8 | 31.7 | 0.15 | **historique** (downgraded -70%, DL/CNN weekly-retraining overfits real fees severely, PSR 3.2%; universe corrected Multi-asset→Equities) |
| 31 | RL-DQN-Trading | RL | Portfolio | ~~0.53~~ → **0.58 (2020-21 only)** ✓post-#2801 | 18.2 | 33.2 | 0.55 | **non re-verifiable** (locked to ~1yr window, runtime error on extension, PSR 30.2%) |
| 31b | RL-Portfolio-Q-Learning | RL | Equities | 0.58 | 18.2 | 33.2 | — | historique (2020-2021) |
| 32 | LSTM-Forecasting | DL | Multi-asset (SPY/QQQ/IWM/EFA/TLT/GLD/IEF) | ~~0.53~~ → **0.525** ✓post-#2801 | 11.3 | 32.5 | 0.35 | robuste (confirmed flat -1%, near-immune despite weekly rebalance — fee-homogeneous 7-US-ETF basket + concentrated 2-4 selection = low realized turnover; PSR 13.4% low not a true leader; sklearn MLPClassifier not real LSTM per docstring) |
| 33 | TrendStocks-Alpha | IND | Equities | ~~0.52~~ → **0.51** ✓post-#2801 | 15.7 | 39.6 | 0.40 | robuste (confirmed -2%, high-turnover near-immune to fees, PSR 5.6%) |
| 34 | Portfolio-IBKR-Coinbase-Hybrid | COMP | Multi-asset (5 IBKR equity sleeves + 3 Binance crypto sleeves) | 0.52 → **1.153** ✓post-#2801 | 38.9 | 38.8 | 1.00 | robuste (**TRUE LEADER PSR 61.95%** — heavy bull-market caveat; catalog 0.52 was a short 383-date baseline) |
| 35 | Framework_Composite_FamaFrenchAllWeather | COMP | Multi-asset (VLUE/MTUM/SIZE/QUAL/USMV + SPY/IEF/GLD/XLP) | — → **0.684** ✓post-#2801 | 13.7 | 7.0 | 1.95 | robuste (**TRUE LEADER PSR 87.5%** — gap-fill first real data; 20/80 FamaFrench/AllWeather composite, monthly rebalance on fee-homogeneous 9-ETF basket HOLDS where AllWeather standalone collapsed -30%; OOS 2023-2026, MaxDD -7% exceptional) |
| 36 | Framework_Composite_EMATrend | COMP | Equities (AAPL/MSFT/GOOGL/AMZN/NVDA + JPM/V/MA/UNH/JNJ/XOM/CVX/HD/PG/KO) | — → **0.741** ✓post-#2801 | 18.7 | 28.0 | 0.67 | robuste (gap-fill first real data; docstring claimed 0.867 → real 0.741 -14%; 70/30 EMA-Cross/TrendStocks composite, weekly rebalance on fee-homogeneous 15 mega-caps; PSR 27.4% not a true leader; Mag7-heavy EMA sleeve = survivorship caveat) |
| 37 | composite-c1-multiasset | COMP | Multi-asset (SPY/TLT/GLD/USO/EFA) | — → **0.175** ✓post-#2801 | 4.7 | 17.0 | 0.28 | **historique** |
| 38 | composite-c2-equityfactor | COMP | Equities (top-25 mkt-cap US) | — → **0.543** ✓post-#2801 | 10.0 | 18.8 | 0.53 | robuste (gap-fill first real data; just above threshold, PSR 16.8% low not a true leader) |
| 39 | HAR-RV-Kelly | RISK | Multi-asset | — → **0.75** ✓post-#2801 | 23.0 | 48.3 | 0.48 | robuste borderline (gap-fill first real data, PSR 24.0%, MaxDD -48% crypto tail) |

### Post-#2801 verification — findings (2026-06-15)

Twenty-one Tier-1 entries re-run live via MCP qc-mcp (project native period, IBKR margin
brokerage = the #2801 Lot 1 remediation). Results vs the pre-remediation catalog values:

| Strategy | QC project | Catalog Sharpe | **Verified Sharpe** | Delta | Real status |
|----------|-----------|---------------|---------------------|-------|-------------|
| HighBookToMarketFScore | 32732820 | 2.09 | **0.41** | -80% | **historique** (was robuste) |
| PuppiesOfTheDow | 32732704 | 1.99 | **0.30** | -85% | **historique** (was robuste) |
| LeveragedETFMomentum | 32732756 | 1.80 | **1.78** | -1% | robuste (confirmed, leveraged) |
| Positive-Negative-Splits-ML | 30317350 | 1.74 | **1.51** | -13% | **non robuste** (revu par #19242 : reproduction 2018 → 2024-04 Sharpe 1,16 / PSR 43,6 %, 89 % du résultat en 2023-2024, capacité estimée par QC 3 000 dollars ; 2018 → 2026-09 : ruine en vente à découvert, variante long seul `NO BEATS`) |
| MacroFactorRotation | 32730301 | 1.23 | **0.73** | -40% | robuste (revised Sharpe) |
| Framework_Composite_TrendWeather | 28825740 | 1.16 | **1.14** | -2% | robuste (confirmed, **PSR 77.9%**) |
| Trend-Following | 28797562 | 1.07 | **0.41** | -62% | **historique** (was robuste) |
| Multi-Layer-EMA | 28433748 | 0.93 | **0.80** | -14% | robuste (confirmed, **ML crypto holds**, PSR 23.9%, MaxDD -57%) |
| EMA-Cross-Stocks | 28789946 | 0.87 | **0.99** | +14% | robuste (confirmed, PSR 49.7%) |
| AllWeather | 28657833 | 0.67 | **0.47** | -30% | **historique** (was robuste — low-turnover multi-asset NOT immune) |
| VolTarget-Momentum | 30784745 | 0.65 | **0.50** | -23% | robuste borderline (PSR 9.4%, at threshold) |
| Crypto-MultiCanal | 30750734 | 0.58 | **0.33** | -43% | **historique** (was robuste — crypto indicator NOT robust post-fees) |
| MomentumStrategy | 28657837 | 0.57 | **0.50** | -12% | robuste borderline (at threshold, PSR 9.3%) |
| TrendStocks-Alpha | 28885507 | 0.52 | **0.51** | -2% | robuste (confirmed, high-turnover near-immune) |
| ML-Temporal-CNN | 29443034 | 0.73 | **0.46** | -37% | **historique** (DL/CNN overfits real fees, PSR 5.2%) |
| HAR-RV-Kelly | 31650567 | — | **0.75** | (gap-fill) | robuste borderline (first real data, PSR 24.0%, MaxDD -48% crypto) |
| RL-DQN-Trading | 32057969 | 0.53 | **0.58 (2020-21)** | n/a | **non re-verifiable** (~1yr window, runtime error on extension) |
| MeanReversion | 30776121 | 0.81 | **0.81** | 0% | robuste (confirmed, PSR 46.8%, low-turnover sector rotation holds) |
| SectorMomentum | 29686886 | 0.62 | **0.56** | -9% | robuste (mild — monthly rebalance winner-takes-all on 3 ETFs = low rotation near-immune; PSR 13.8% low, not a true leader) |
| BlackLitterman-Momentum | 29816300 | 0.60 | **0.83** | +38% | **robuste (catalog CORRECTED UPWARD — was stale, not a fee effect; code already remediated, real Sharpe matches docstring v2 BEST 0.823; monthly fee-homogeneous 15-stock US equity = near-immune; PSR 51.4% true leader borderline)** |
| ML-FeatureEngineering | 29808616 | 0.57 | **0.65** | +15% | robuste (catalog CORRECTED UPWARD — was stale; current RF+GB 18-feature ensemble better than the run that produced 0.57; bi-weekly fee-homogeneous US equity = near-immune; PSR 15.1% low, not a true leader; universe corrected Multi-asset→Equities; baseline-clone 32952140) |
| LongShortHarvest | 32921183 | 3.39 | **1.64** | -52% | robuste (confirmed, **PSR 98.7%** top — catalog was 1Y-OOS, baseline-clone) |
| DynamicVIXSpyRegime | 32921262 | 1.72 | **1.00** | -42% | robuste (confirmed, **PSR 69.4%** — catalog was 1Y-OOS, baseline-clone) |
| Portfolio-Optimization-ML | 29318874 | 0.90 | **0.88** | -2% | robuste (confirmed, monthly-rebalance fee-homogeneous equity basket near-immune, PSR 37.2%) |
| CausalEventAlpha | 29809163 | 0.78 | **0.45** | -43% | **historique** (was robuste — monthly cadence but full-basket sector rotation = high turnover, PSR 5.5%) |
| Gaussian-Direction-Classifier | 29398513 | 0.76 | **0.76** | 0% | robuste (confirmed — daily schedule but low realized turnover on liquid mega-cap = near-immune, PSR 22.7%) |
| ML-Trend-Scanning | 29808859 | 0.66 | **0.33** | -50% | **historique** (was robuste — daily-rebalance SPY/TLT/GLD multi-asset crushed by real fees, PSR 7.8%) |
| TrendStocksLite | 28817425 | 0.72 | **0.71** | -2% | robuste (confirmed — weekly trend on 15 liquid large-caps, low realized turnover, near-immune, PSR 25.0%) |
| ML-RandomForest | 29434751 | 0.68 | **0.70** | +3% | robuste (confirmed — bi-weekly RF on 10 mega-caps, moderate turnover near-immune on fee-homogeneous US equity, PSR 18.4%, baseline-clone 32940005) |
| ML-XGBoost | 29434753 | 0.57 | **0.555** | -3% | robuste (confirmed flat-to-mildly-down, NOT an upward correction — bi-weekly GradientBoostingRegressor on fee-homogeneous 15 mega-caps = near-immune; PSR 10.6% low, not a true leader; universe corrected Multi-asset→Equities; baseline-clone 32958201) |
| RegimeSwitching | 28693650 | 0.55 | **0.581** | +6% | robuste (confirmed — multi-asset HOLDS via explicit turnover suppression: regime-change-only rebalance + anti-micro-rebalancing delta<5% + beta-annealing; contrasts AllWeather -30%; refines discriminator = realized-turnover not asset-class; PSR 7.0% low; universe Equities/ETF→Multi-asset; baseline-clone 32961367) |
| Markov-Regime-Detection | 29398512 | 0.57 | **0.375** | -34% | **historique** (downgraded — catalog 0.57 was STALE; docstring v1.0 claimed 0.408; IBKR fees only ~8% drag from intrinsic ~0.41, NOT a fee-collapse. CHALLENGES discriminator: same multi-asset (SPY/TLT/GLD) + turnover-suppression as RegimeSwitching (+6% held) yet collapses; difference = per-event trade SIZE — Markov does binary 80% SPY↔TLT swap per regime change (high per-trade impact despite low freq) vs RegimeSwitching partial 70/30 + persistent defensive sleeve. Discriminator v3 = frequency × per-trade-size; universe Equities→Multi-asset; baseline-clone 32964364) |
| Temporal-CNN-Prediction | 29816576 | 0.54 | **0.161** | -70% | **historique** (downgraded — DL/CNN weekly-retraining overfits real fees severely; same regime as ML-Temporal-CNN 0.73→0.46; PSR 3.2%; universe corrected Multi-asset→Equities (QQQ top-3); baseline-clone 32967763 + temporalcnn.py lib copy) |
| LSTM-Forecasting | 29443476 | 0.53 | **0.525** | ~0% | robuste (confirmed near-flat — multi-asset 7-US-ETF (SPY/QQQ/IWM/EFA/TLT/GLD/IEF) weekly-rebalance NN HOLDS; fee-homogeneous ETF basket + concentrated 2-4 selection = low realized turnover (discriminator v3); contrasts Temporal-CNN -70% DL collapse; sklearn MLPClassifier (64,32) not real LSTM per docstring; PSR 13.4% low; baseline-clone 32970990) |
| Framework_Composite_FamaFrenchAllWeather | 28882145 | — | **0.684** | gap-fill | robuste (**TRUE LEADER PSR 87.5%** — first real backtest; 20% FamaFrench (VLUE/MTUM/SIZE/QUAL/USMV factor ETFs) + 80% AllWeather (SPY/IEF/GLD/XLP) composite, monthly rebalance on fee-homogeneous 9-US-ETF basket HOLDS where AllWeather standalone collapsed -30% (0.67→0.47); FamaFrench factor diversification + monthly low-frequency = low realized turnover; already had IBKR (direct backtest, no clone); OOS 2023-2026, Calmar 1.95, MaxDD -7% exceptional) |
| Framework_Composite_EMATrend | 28911253 | — | **0.741** | gap-fill | robuste (first real backtest — docstring claimed 0.867 → real 0.741, -14% from claim; 70% EMA-Cross (AAPL/MSFT/GOOGL/AMZN/NVDA Mag7) + 30% TrendStocks (15 mega-caps) composite, weekly rebalance on fee-homogeneous US equity basket; already had IBKR (direct backtest, no clone); decade 2015-2025, PSR 27.4% not a true leader; **Mag7 survivorship caveat** — EMA sleeve 100% Mag7, decade dominated by Mag7 outperformance inflates trend signal) |
| composite-c1-multiasset | 32981093 | — | **0.175** | gap-fill | **historique** (first real backtest — local-only framework deployed to QC Cloud; 5-ETF multi-asset rotation SPY/TLT/GLD/USO/EFA, 3-alpha ensemble (Momentum/MACD/RelativeStrength), RiskParityPCM weekly rebalance, drawdown-cap 12% + trail 4% + VWAP execution; multi-asset rotation through USO (commodity) + EFA (intl equity) = friction/volatility basket churns, same regime as AllWeather -30% collapse; PSR 2.0% very low; framework already had IBKR so direct deploy no clone; totalOrders=0 = phantom MCP bug, CAGR 4.7% proves trades executed) |
| composite-c2-equityfactor | 32981222 | — | **0.543** | gap-fill | robuste (first real backtest — local-only framework deployed to QC Cloud; 25-stock equity factor composite, FineFundamental top-25 mkt-cap, 4-factor ensemble (Value/Quality/LowVol/Momentum), MeanVariancePCM weekly, sector-cap 18% + portfolio DD 18% + TWAP execution; fee-homogeneous mega-cap US equity basket HOLDS just above threshold; **contrasts c1 historique** — same composite architecture but c2 equity-only fee-homogeneous holds while c1 multi-asset friction basket collapses, re-confirming discriminator v3 (realized turnover × fee-homogeneity); PSR 16.8% low not a true leader; IBKR already in framework, direct deploy no clone; totalOrders=0 = phantom MCP bug, CAGR 10.0% proves trades) |
| Portfolio-IBKR-Coinbase-Hybrid | 31717642 | 0.52 | **1.153** | n/a (period) | robuste (**TRUE LEADER PSR 61.95%**, heavy caveat) (first FULL-period backtest — catalog 0.52 was a **short 383-date baseline mislabeled "2020-2025"** [CAGR 15.7%/MaxDD 16.9%, only ~1.5y]; current Phase-3 paramétré version on FULL 2020-2025 [1828 dates] = Sharpe 1.153/CAGR 38.9%/MaxDD 38.8%, cross-checked on 2023-2025 OOS 50/50 = 1.321/PSR 77.98%; **BULL-MARKET CAVEAT** ~40% CAGR is NOT sustainable — 50% crypto sleeve + 50% equity rode the 2020-21 + 2023-24 crypto/Mag7 bull, sample lacks the 2018 crypto winter or any prolonged bear → robuste on bullish years but NOT regime-robust; **multi-broker caveat** — IBKR rejects Crypto (#1027 P2) so no single set_brokerage_model(IBKR); uses explicit PercentFeeModel 5bps equity/10bps crypto + 5bps slippage = cost-realistic [the #2801 spirit], the strategy was ALWAYS cost-realistic so catalog 0.52 was NOT fee-optimistic, just a truncated run; project = #1027 main.py Phase-3, backtest via params start/end/ibkr_alloc, no code change) |

**Finding (methodological, now 37-strategy sample)** : the remediation impact is **not
uniform**, and the batch-4 results *refine and partly correct* the earlier 10-strategy pattern.
The distinguishing axis is **not** asset class, nor ML-vs-indicator alone — it is the
combination of (a) the fee-per-trade the asset class carries and (b) how the strategy turns
over against that fee. Four regimes now observed:

- **Value/factor + simple trend COLLAPSE** (-62% to -85%): HighBookToMarketFScore, PuppiesOfTheDow,
  Trend-Following. These top-ranked catalog entries drop out of "robuste" entirely — their
  pre-remediation Sharpes were artifacts of the negligible default fee model. MacroFactorRotation
  (-40%) is the milder case in this family. **These catalog rankings are not reliable.**

- **Structured ML & regime composites HOLD** (-2% to -14%): Positive-Negative-Splits-ML (1.51,
  **PSR 82.3%**, statut revu par #19242 : non robuste), Framework_Composite_TrendWeather (1.14, **PSR 77.9%**), and Multi-Layer-EMA
  (0.80, PSR 23.9%) confirm. Contrast with MomentumRegime (0.185) and now Crypto-MultiCanal —
  only the *structured* ML/regime-aware designs survive real fees. **These are real alpha.**

- **High-turnover US equity is near-immune** (0% to -2%): EMA-Cross-Stocks, TrendStocks-Alpha, and
  now TrendStocksLite (0.72→0.71, -2%, weekly trend on 15 liquid large-caps — slow SMA200/EMA signals,
  low realized turnover) barely move under real IBKR fees, confirming the #1407 finding that US equity fees (<0.25 bps/trade)
  are negligible even at high turnover.

  **Batch 5 refines this: the immunity is signal-FREQUENCY,
  not asset-class — ML-Temporal-CNN (QQQ equity, -37% to 0.46, DL signal-churning) erodes despite
  being US equity. Slow EMA/trend signals (few trades) are immune; DL/CNN direction predictions
  (frequent re-entry) are not.

  Batch 6 confirms: MeanReversion (low-turnover sector rotation,
  0.81→0.81, **0% delta**, PSR 46.8%) holds flat — it is multi-asset like AllWeather but, unlike
  AllWeather, rotates *within a single fee-homogeneous equity class* (<0.25 bps/trade), so its slow
  signals incur negligible cost. AllWeather's -30% drop came from its bonds/gold sleeve crossing
  higher per-trade friction. The discriminator is thus **signal-frequency × fee-homogeneity of the
  traded basket**, not asset-class alone.**

  Batch 7 sharpens this further: cadence is not turnover.
  Portfolio-Optimization-ML (monthly, fee-homogeneous 15-stock basket, **low** turnover — covariance
  weights nudge rather than rotate) is near-immune (0.90→0.88, -2%), whereas CausalEventAlpha (also
  monthly, also a fee-homogeneous 8-sector-ETF basket) **collapses** (0.78→0.45, -43%, PSR 5.5%) because
  it liquidates and re-ranks the entire basket each month and concentrates to 4 sectors in bear regimes
  — high per-period turnover. A monthly schedule over a homogeneous basket is therefore *not* sufficient
  for immunity; the driver is realized turnover per rebalance. Gaussian-Direction-Classifier closes
  the proof from the other side: a *daily* schedule (higher cadence than either) is near-immune
  (0.76→0.76, 0%) because its realized turnover is low — max-3 positions whose mega-cap probability
  rankings rarely flip, over an ultra-liquid basket. Higher cadence with low realized turnover beats
  lower cadence with full rotation; cadence alone predicts nothing.

- **Low-turnover multi-asset & crypto indicators are NOT immune** (-30% to -43%): batch 4
  *invalidates* the earlier broad "ML/crypto holds" generalization. AllWeather (low-turnover
  multi-asset, -30% → 0.47) and Crypto-MultiCanal (crypto indicator, -43% → 0.33) both drop
  below the robuste threshold. Crypto's 10bps fees + the indicator's signal-chasing turnover erode
  it hard; the Binance CASH cash-constraint benefit seen in the #1407 fee sweep (0.181→0.333)
  still leaves it well under the catalog 0.58. HAR-RV-Kelly (vol-targeting Kelly, 0.75, PSR 24.0%)
  is the exception that proves the rule: Kelly position-sizing dampens exposure, so it survives fees
  where the equal-weight indicator (Crypto-MultiCanal, 0.33) collapses — but its -48% drawdown and
  non-significant PSR mark it as a risk overlay, not pure alpha.

**Reproducibility failure mode (distinct from fee-collapse)** : RL-DQN-Trading (catalog 0.53)
cannot be re-verified on the remediation window — it runs only on a ~1-year slice (253 tradeable
dates, 2020-2021; Sharpe 0.58, PSR 30.2%) and Runtime-Errors on any date extension. This is not a
fee effect; it is RL training-window lock-in. The catalog "0.53" is real but un-generalizable.

**Borderline band** (-12% to -23%, sitting on the 0.5 line): VolTarget-Momentum (0.50, PSR 9.4%)
and MomentumStrategy (0.50, PSR 9.3%) — both non-significant PSR, technically robuste but on the edge.

**True leaders post-#2801, ranked by statistical significance (PSR > 50%)** :

1. Positive-Negative-Splits-ML — 1.51, PSR 82.3% (structured ML, top leader, replaces collapsed value entries) — **statut retiré par #19242** : la mesure ne se reproduit pas (Sharpe 1,16, PSR 43,6 % sur 2018 → 2024-04) et la règle se ruine en vente à découvert en décembre 2025
2. LeveragedETFMomentum — 1.78, PSR 79.8% (leveraged, extreme profile)
3. Framework_Composite_TrendWeather — 1.14, PSR 77.9% (regime-aware composite)
4. EMA-Cross-Stocks — 0.99, PSR 49.7% (high-turnover US equity, near-significance, near-immune)

Only these 4 survive both the fee remediation AND a significance bar. The "robuste" Tier-1 band
of 41 catalog entries shrinks to roughly **a dozen genuinely-holding strategies** under real fees;
the rest are overstated to varying degrees.

**Implication for the réunion Nicolas 15/06** : the catalog is not uniformly stale, but the
overstatement is widespread — **only 6 of 24 verified strategies hold robuste with significant PSR**.
The overstatement is structural in two families (value/factor/trend, and crypto indicators), while
structured ML and regime-aware composites are validated. The comparative table MUST be cited by
significance (PSR) not raw Sharpe; collapsed entries need a caveat before any pedagogical use. The
remaining Tier-1 list (19 strategies) needs systematic re-backtest before the table is trusted
end-to-end. LongShortHarvest-QC (catalog 3.39, the single highest entry) and DynamicVIXSpyRegime-QC
(1.72) — previously QC Community Library references without an owned project — were deployed as owned
baseline-clones (project IDs 32921183 / 32921262) and re-run over 2015-2024: both keep robuste status
(PSR 98.7% / 69.4%) but their headline Sharpes nearly halve (1.64 / 1.00), confirming the catalog
1Y-OOS values were optimistic while the strategies themselves are sound.

Backtests: `1630-baseline-HighBookToMarketFScore-post2801` (0.411, 14.5%, -60.4%, PSR 4.5%),
`1630-baseline-PuppiesOfTheDow-post2801` (0.302, 9.6%, -28.8%, PSR 3.5%),
`1630-baseline-LeveragedETFMomentum-post2801` (1.779, 126.4%, -53.3%, PSR 79.8%),
`1630-baseline-PositiveNegativeSplits-post2801` (1.511, 75.7%, -37.6%, PSR 82.3%),
`1630-baseline-MacroFactorRotation-post2801` (0.731, 22.6%, -42.0%, PSR 23.8%),
`1630-baseline-TrendWeather-post2801` (1.14, 27.1%, -27.7%, PSR 77.9%),
`1630-baseline-TrendFollowing-post2801` (0.407, 7.89%, -14.6%, PSR 8.7%),
`1630-baseline-EMACrossStocks-post2801` (0.991, 29.2%, -35.7%, PSR 49.7%),
`1630-baseline-VolTargetMomentum-post2801` (0.50, 11.1%, -21.2%, PSR 9.4%),
`1630-baseline-TrendStocksAlpha-post2801` (0.512, 15.7%, -39.6%, PSR 5.6%),
`1630-baseline-MultiLayerEMA-post2801` (0.798, 25.0%, -57.1%, PSR 23.9%),
`1630-baseline-AllWeather-post2801` (0.469, 7.5%, -16.4%, PSR 19.6%),
`1630-baseline-CryptoMultiCanal-post2801` (0.333, 4.6%, -14.1%, PSR 13.0%),
`1630-baseline-MomentumStrategy-post2801` (0.499, 11.2%, -25.8%, PSR 9.3%),
`1630-baseline-LongShortHarvest-post2801` (1.635, 45.6%, -17.0%, PSR 98.7%),
`1630-baseline-DynamicVIXSpyRegime-post2801` (0.997, 17.9%, -16.5%, PSR 69.4%),
`1630-PortfolioOptimizationML-post2801` (0.884, 27.2%, -41.6%, PSR 37.2%),
`1630-CausalEventAlpha-post2801` (0.447, 11.7%, -38.8%, PSR 5.5%),
`1630-GaussianDirectionClassifier-post2801` (0.761, 23.1%, -25.6%, PSR 22.7%),
`1630-TrendStocksLite-post2801` (0.707, 18.0%, -33.7%, PSR 25.0%),
`1630-ML-RandomForest-post2801` (0.70, 20.6%, -40.9%, PSR 18.4%).

---

## Tier 2 — Historique (Sharpe 0.0-0.5, 17 strategies)

Backtested but modest or marginal strategies. Useful for pedagogical comparison.

| # | Project | Type | Class | Sharpe | CAGR% | MaxDD% | Calmar | Status |
|---|---------|------|-------|--------|-------|--------|--------|--------|
| 40 | Sector-ML-Classification | ML | Equities (sectors) | 0.47 | — | — | — | historique |
| 41 | EMA-Cross-Index | IND | Equities (SPY) | 0.47 | — | — | — | historique |
| 42 | DualMomentumNoTLT | IND | Multi-asset | 0.47 | — | — | — | historique |
| 43 | Dividend-Harvesting-ML | ML | Equities | 0.47 | — | — | — | historique |
| 44 | Adaptive-Conformal-Risk | ML | Multi-factor (15 US large-caps) | ~~0.42~~ → **0.449** ✓post-#2801 | 11.5 | 22.5 | 0.51 | **historique** (first real #1630-aligned vetting — catalog 0.42 unverified; monthly rebalance + ACI vol-targeting = low turnover → modest fee drop confirms #1630 discriminator; PSR 12.5% non-significant; **Mag7-survivorship caveat** universe AAPL/MSFT/NVDA/AMZN/TSLA 5/7 Mag7 → Sharpe survivorship-driven not ACI value; real ACI Gibbs-Candès 2021 but can't isolate on biased universe; cf EMATrend #36 same caveat class; see finding #38) |
| 45 | PCA-StatArbitrage | ML | Equities (top-100) | ~~0.40~~ → **0.205** ✓post-#2801 | 6.76 | 31.8 | 0.21 | **historique** (first real #1630-aligned vetting — catalog 0.40 does NOT survive: 0.40→0.205 PSR 1.4% noise-level; **fee-IMMUNE proven** — hardcoded no-brokerage clone byte-identical 0.205, so the drop is WINDOW-effect not fee, the aligned 2024-25 Mag7-momentum phase hostile to contrarian PCA mean-reversion; monthly full-rotation predicted fee-vulnerable but fee-homogeneity of US-equity basket dominates → confirms #1630 discriminator regime 3 near-immune, cf EMA-CS IBKR 0.991=nobrokerage; see finding #40) |
| 46 | RiskParity | RISK | Multi-asset | ~~0.40~~ → **0.361** ✓post-#2801 | 8.3 | 18.4 | 0.45 | **historique** (first real #1630-aligned vetting — catalog 0.40 unverified; true inverse-volatility risk-parity, rank-based RP v4 "BEST" docstring 0.515, on SPY/EFA/GLD/DBC/TLT + BND safe-haven, monthly, IBKR margin; aligned 0.361 = -30% vs student 2015-2025 0.514 proj 31872286 *non-aligned = regime-4 fee-friction, bonds+commodities NOT fee-homogeneous = AllWeather -30% pattern; PSR 11.3% non-significant; best risk-parity-family member cf Cloud-RiskParity-Composite 0.027 #23; cf residual-pool subsection) |
| 47 | ML-Gaussian-Classifier | ML | Equities | 0.36 | — | — | — | historique |
| 48 | DualMomentum | IND | Multi-asset | ~~0.35~~ → **0.350** ✓post-#2801 | 8.4 | 14.9 | 0.56 | **historique** (first real #1630-aligned vetting — catalog 0.35 confirmed essentially FLAT, no fee-collapse; Antonacci canonical dual momentum, absolute filter price>SMA200 AND 6m>0 + relative 12m ranking all-candidates momentum-tilt, on SPY/EFA/EEM/TLT/GLD/DBC + BND safe-haven, monthly, IBKR margin; aligned 0.350/MaxDD 14.9% = tighter drawdown than sibling RiskParity 18.4% (dual-filter + BND refuge); PSR 9.6% non-significant — student 0.493/PSR 54.9 proj 31798582 was a 2-yr small-sample artifact = the FamaFrenchAllWeather 87.5%→22.9% pattern finding #35; cf residual-pool subsection) |
| 49 | ML-Reversion-Trending | ML | Multi-asset | 0.29 | — | — | — | historique |
| 50 | BTC-ML | ML | Crypto (BTC) | 0.28 | — | — | — | historique |
| 51 | ML-Chronos-Foundation | DL | Multi-asset | 0.28 | — | — | historique |
| 52 | Chronos-Foundation-Forecasting | ML | Multi-asset | 0.25 | — | — | historique |
| 53 | EMA-Cross-Crypto | IND | Crypto (BTC) | 0.24 | — | — | historique |
| 54 | InverseVolatility-Rank | ML | Futures | 0.21 | — | — | historique |
| 55 | OptionsIncome | OPT | Options (SPY) | 0.21 | — | — | historique |
| 56 | Framework_Composite_MomentumRegime | COMP | Multi-asset | 0.19 | 4.7 | — | historique |

---

## Tier 3 — Exploratoire (Sharpe < 0, 5 strategies)

Negative Sharpe — either failed strategies or market conditions unfavorable.

| # | Project | Type | Class | Sharpe | CAGR% | Status |
|---|---------|------|-------|--------|-------|--------|
| 58 | EMA-Cross-Alpha | IND | Equities | -0.01 | 2.8 | exploratoire |
| 59 | TrendFilteredMeanReversion | IND | Equities (SPY) | -0.02 | — | exploratoire |
| 60 | ForexCarry | IND | FX | -1.11 | -0.5 | exploratoire |
| 61 | PairsTrading | STAT | Equities | -0.36 | — | exploratoire |
| 62 | ETF-Pairs | STAT | ETF | -0.71 | — | exploratoire |

---

## Tier 4 — Untested (candidates for standardized backtest, ~38 strategies)

Projects with `main.py` but no recorded backtest metrics. Prime candidates for the #1630 baseline run.

| # | Project | Type | Class | research.ipynb |
|---|---------|------|-------|----------------|
| 63 | ML-Classification | ML | Equities | yes |
| 64 | ML-Regression | ML | Equities | yes |
| 65 | ML-Ensemble | ML | Equities | yes |
| 66 | ML-EnhancedPairs | ML | Equities | yes |
| 67 | ML-DeepLearning | DL | Equities | yes |
| 68 | DL-LSTM | DL | Equities | yes |
| 69 | ML-TextClassification | ML/NLP | Equities | yes |
| 70 | RL-Portfolio | RL | Multi-asset | yes |
| 71 | Reinforcement-Learning-Trading | RL | — | yes |
| 73 | Option-Wheel | OPT | Options (SPY) | yes |
| 75 | Options-VGT | OPT | Options (VGT) | yes |
| 76 | Crypto-LSTM-Prediction | DL | Crypto (BTC) | yes |
| 92 | Research-Executor | — | Multi-asset | yes |

> **Tier-4 residual characterization (2026-06-23, #1630)**: the no-ML IND/COMP/RISK/FACTOR candidates are all verified (Key-findings #22–#36). The remaining Tier-4 rows are: **ML/DL/RL** (#63–#71, #76) — heavy training, deferred to the training-specialist and to multi-seed cross-validation (Next Steps #6); **OPT** (#73 Option-Wheel, #75 Options-VGT) — naked-options strategies whose MaxDD exceeds 100 % (the QC simulator does not capture forced assignment/liquidation, see the caveat under Key-finding #13), so the baseline is intrinsically optimistic on loss and of low pedagogical value (the student OptionWheel Sharpe −0.51 / MaxDD 103.5 % already documents the catastrophe); **#92 Research-Executor** — a *research execution harness* (runs 8 embedded notebooks via a `MockQB` shim over a 2-day window and `quit()`s inside `initialize`, `on_data = pass`), **not a tradable strategy**, hence out of the #1630 aligned-baseline scope (see Key-finding #37).

---

## Type Distribution

| Type | Count | Avg Sharpe (tested) |
|------|-------|---------------------|
| ML | 33 | 0.62 |
| IND (indicator) | 17 | 0.55 |
| COMP (composite) | 12 | 0.65 |
| DL (deep learning) | 7 | 0.48 |
| RISK (risk-parity/vol) | 9 | 0.57 |
| RL (reinforcement) | 4 | 0.53 |
| OPT (options) | 4 | 0.17 |
| STAT (stat-arb) | 2 | -0.54 |
| FACTOR | 1 | — |

---

## Asset Class Distribution

| Asset Class | Count |
|-------------|-------|
| Multi-asset | 35 |
| Equities | 42 |
| Crypto (BTC) | 7 |
| Options | 4 |
| FX | 2 |
| Futures/Commodities | 2 |

---

## #1630 Aligned Baselines (2018-2025 period)

Standardized backtest results from QC Cloud via MCP qc-mcp-lite. Period: 2018-01-01 to 2025-12-31 (US equities/multi-asset), 2020-01-01 to 2025-12-31 (crypto). Some strategies have hardcoded dates that cannot be changed without breaking ML logic.

### Verified baselines (CAGR/MaxDD from QC Cloud API)

| Project | QC ID | Period | Sharpe | CAGR% | MaxDD% | PSR% | Backtest ID | Notes |
|---------|-------|--------|--------|-------|--------|------|-------------|-------|
| TrendFollowing | 28797562 | 2018-2025 | **1.072** | 23.2 | 9.3 | 81.8 | `7792ae0a` | Leader, PSR > 80% |
| AllWeather | 28657833 | 2010-2025* | 0.631 | 9.016 | 16.4 | 31.2 | `cd6ba790` | *Hardcoded start 2010. PSR 31% |
| SectorDualMomentum | 29686886 | 2015-2025* | 0.581 | 13.488 | 22.8 | 15.3 | `8713e974` | *Hardcoded start 2015 |
| MomentumStrategy (SectorMom) | 28657837 | 2010-2025* | 0.555 | 11.676 | 25.8 | 7.2 | `e4491127` | *Hardcoded start 2010 |
| VolTarget-Momentum | 30784745 | 2018-2025 | 0.648 | 14.7 | 21.2 | 22.3 | `c3223fe5` | Confirmed |
| Crypto-MultiCanal | 30750734 | 2020-2025 | 0.581 | 8.2 | 17.0 | 37.6 | `4e97d7dc` | Stable crypto |
| EMA-Cross-Stocks | 28789946 | 2018-2025 | **0.891** | 26.229 | 35.7 | 40.5 | `6b40d921` | Highest CAGR |
| Portfolio-IBKR-Binance | 31717642 | 2020-2025 | 0.519 | 15.7 | 16.9 | 46.2 | `4cbb9476` | Multi-asset |
| TrendStocks-Alpha | 28885507 | 2018-2025 | 0.519 | 15.9 | 39.6 | 5.8 | `7c434dbd` | High MaxDD |
| EMA-Cross-Alpha | 28885488 | 2018-2025 | -0.010 | 2.8 | 14.0 | 0.5 | `633779d0` | Period overfitting |
| MomentumRegime | 31243821 | 2018-2025 | 0.185 | 4.7 | 11.5 | 13.0 | `033834d8` | Double-defense |
| ForexCarry | 28657908 | 2015-2025* | -1.108 | -0.5 | 19.2 | — | `c3afe374` | *Cannot restrict to 2018+ |
| RL-Q-Learning | 32057969 | 2020-2021* | 0.584 | 18.2 | 33.2 | — | `fb1a6366` | *Hardcoded dates |
| PuppiesOfTheDow-QC | 32732704 | 2018-2025 | 0.302 | 9.613 | 28.8 | 3.5 | `37266daa` | Collapse vs catalog 1.99 (period overfitting) |
| LeveragedETFMomentum-QC | 32732756 | 2018-2025 | **1.779** | 126.388 | 53.3 | **79.8** | `2addf467` | Confirms catalog 1.80. Lev ETF: extreme CAGR, MaxDD 53% |
| Framework_Composite_TrendWeather | 28825740 | 2018-2025 | 0.948 | 24.603 | 27.5 | 56.6 | `cd84c50b` | Close to catalog 1.16, best composite |
| HighBookToMarketFScore-QC | 32732820 | 2018-2025 | 0.411 | 14.513 | 60.4 | 4.5 | `5ef58b0d` | Collapse vs catalog 2.09, MaxDD 60% |
| MeanReversion v5.2 | 30776121 | 2015-2024 | **0.810** | 10.040 | 7.5 | 46.8 | `b2c5b08f` | Promoted Tier 2→1. Calmar 1.34 (best risk-adj). PSR 46.8% |
| AdaptiveAssetAllocation | 28693649 | 2008-2024 | 0.509 | 8.008 | 18.9 | 10.6 | `89e8aaef` | Promoted Tier 4→1. Min-var + momentum |
| Cloud-MeanReversion-Sectors (v1) | 33211207 | 2018-2025 | 0.067 | 3.793 | 41.7 | 1.0 | `85bcef2b` | RSI(14) mean-rev 11 sector ETFs, IBKR. v1 default (no regime filter). Near-zero aligned, PSR 1.0% noise, MaxDD 42% (no regime filter exits bears). Promoted Tier 4→2. v2/v3 (regime filter + stop-loss) variants untested |
| PairsTrading | 28693651 | 2015-2024 | -0.280 | 1.101 | 15.9 | 0.0 | `1ed0de9d` | Confirms exploratoire. PSR 0.001% |
| AssetClassMomentum-QC | 33209767 | 2018-2025 | 0.22 | 6.644 | 28.1 | 3.8 | `6746f155` | 5-ETF momentum top-3 (SPY/EFA/BND/VNQ/GSG, 252d), IBKR. Weak aligned, PSR 3.8% non-significant. Promoted Tier 4→2 |
| Cloud-RiskParity-Composite | 30820857 | 2018-2025 | 0.027 | 3.504 | 24.4 | 1.2 | `b184b13e` | 6-asset rotation (SPY/TLT/GLD/EFA/EEM/DBC, SMA200+6m mom, eq-wt, IBKR). Near-flat aligned, PSR 1.2%. Promoted Tier 4→2. totalOrders=0 = wrapper extraction artifact (CAGR 3.5% ⇒ real trades) |
| Cloud-SectorRotation-Momentum | 30821748 | 2018-2025 | -0.029 | 2.125 | 42.7 | 0.5 | `58fb0a94` | 5-ETF momentum-weighted rotation (QQQ/SPY/EFA/GLD/IWM+SHY, SMA200+6m), IBKR. First Tier-3 backbone (negative), momentum-weighting underperformed equal-weight #79. Promoted Tier 4→3 |
| Cloud-VolTargeting (v1) | 30823587 | 2018-2025 | 0.207 | 6.717 | 38.2 | 2.4 | `179ef1c5` | SPY vol-targeting (target_vol 12%, 21d realized-vol, clamp 30-150%, monthly), IBKR. True vol-targeting (unlike the misnomer #79). Promoted Tier 4→2. MaxDD 38% > #77 28%: the 150% upper clamp leveraged the portfolio into vol spikes. totalOrders=0 = wrapper artifact |
| GlobalMacro-Regime (v3) | 30781695 | 2018-2025 | 0.454 | 9.798 | 22.8 | 16.7 | `bbb73b7b` | Rank-based risk-parity + SPY regime switch (Bridgewater-style, SMA200+6m bull/bear, BND/TLT/GLD defensive RP), IBKR. Best backbone baseline yet (0.454 ~ just under Tier 1); regime-switch + inv_vol weighting beats simple momentum rotations (#77/#79/#80) with controlled MaxDD. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| MomentumRegime-AdaptiveWeights (T85/RS15) | 31524424 | 2018-2025 | -0.729 | 1.875 | 4.3 | 17.4 | `6e8f164f` | COMP framework SectorMomentum 85%+RegimeSwitching 15% (QC Alpha/PCM additive; composite momentum 1/3/6/12m + SMA200 filter; RegimeSwitching SPY SMA50/SMA200+RSI), IBKR. Over-defensive: tiny MaxDD 4.3% over 2018-25 (COVID+2022 bear) = mostly IEF/GLD/cash, CAGR 1.875% < risk-free → negative Sharpe. Promoted Tier 4→3. WORSE than baseline 60/40 composite 0.185 = double-defense (Key-finding #27). totalOrders=0 = wrapper artifact |
| TermStructureCommodities-QC | 33224097 | 2018-2025 | -0.244 | -31.478 | 96.8 | 0.007 | `e9f7e686` | IND Commodities no-ML long-short futures on roll returns/backwardation-contango (top-quintile, monthly, 21 commodity futures Softs/Grains/Meats/Energies/Metals), IBKR margin. Catastrophic: CAGR -31.5% / MaxDD 96.8% = nearly-to-zero. Confirms published library header (OOS 5Y Sharpe -0.041, MaxDD 80.8%): roll-yield signal broke through 2020 COVID oil crash + 2022 energy/inflation term-structure dislocation. Promoted Tier 4→3 (3rd Tier-3, worst by CAGR/MaxDD). totalOrders=0 = wrapper artifact |
| HAR-RV-J-Kelly | 31650567 | 2018-2025 | 0.524 | 14.077 | 37.1 | 10.7 | `df687834` | RISK crypto HAR-RV-J vol-forecasting (Corsi 2009 + ABD 2007 jump via Huang-Tauchen BPV, inline OLS np.linalg.lstsq refit 22d, iterated 5d log-RV forecast, 1/4-Kelly × 5d-momentum direction, BTC/ETH/LTC/BCH USDT, Binance). Sharpe 0.524 = best backbone since GlobalMacro-Regime 0.454; first RISK-family signal surviving alignment positive (cf TermStructureCommodities -0.244 #28). MaxDD 37.1% < 2020-25 window 48.3% (longer coef path through 2022 bear = less leverage). PSR 10.7% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| Vol-GARCH-Target | 33245149 | 2018-2025 | 0.325 | 6.971 | 10.8 | 14.9 | `a0f7a5b5` | RISK GARCH(1,1) vol-targeting (variance-targeting MLE inline alpha/beta grid search Gaussian loglik, pure numpy no arch, refit 22d on 500d window; 10% ann vol target, 30% per-asset cap, SMA200 direction, weekly Monday; 6 ETFs SPY/EFA/EEM/TLT/GLD/DBC, IBKR). Sharpe 0.325 positive; MaxDD 10.8% = tightest backbone baseline (cf GlobalMacro-Regime 22.8%, HAR-RV-J-Kelly 37.1%) = genuine risk budgeting; beats single-asset Cloud-VolTargeting #25 (0.207/MaxDD 38.2%) on Sharpe AND MaxDD. PSR 14.9% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| Vol-Ensemble-Conservative | 33248352 | 2018-2025 | 0.265 | 6.142 | 10.4 | 13.6 | `db77ae49` | RISK ensemble GARCH(1,1)+HAR(1,5,22) vol-forecasting (max-of conservative; GARCH variance-targeting MLE inline alpha/beta grid + HAR OLS np.linalg.lstsq, pure numpy; 8% ann vol target half-allocation, 25% per-asset cap, SMA200 regime -50% bear, SMA50 direction, weekly Monday; 6 ETFs SPY/EFA/EEM/TLT/GLD/DBC, IBKR). Sharpe 0.265 positive; MaxDD 10.4% = NEW tightest backbone baseline (beats Vol-GARCH-Target #30 10.8%, GlobalMacro-Regime 22.8%, HAR-RV-J-Kelly 37.1%); but Sharpe BELOW single-GARCH Vol-GARCH-Target 0.325 = conservative overlay (max-ensemble + half-sizing + regime filter) tightened drawdown at cost of return, HAR added no risk-adjusted value over GARCH alone. PSR 13.6% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| FamaFrench | 33251801 | 2018-2025 | 0.445 | 11.111 | 24.1 | 11.9 | `697e96af` | FACTOR risk-adjusted momentum ETF rotation (FactorETFRotation v3.0; 12m-1m return / 63d realized-vol momentum-Sharpe score, dynamic top_n = all-positive, SMA200 SPY regime, USMV risk-off, per-position -12% stop-loss, monthly rebalance; VLUE/MTUM/SIZE/QUAL/USMV, IBKR). Sharpe 0.445 positive = 3rd-best no-ML backbone (close to GlobalMacro-Regime 0.454 #26); CAGR 11.1% = 2nd-highest (after HAR-RV-J-Kelly 14.1% #29). Mild Sharpe drop vs author v3.0 2015-2024 (0.540->0.445; MaxDD stable 24.2->24.1%) = robust, not period-overfit, contrasting the "simple factor collapse" (this is risk-adjusted momentum rotation, not static factor). Below FamaFrenchAllWeather composite 0.684 = framework adds value. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| composite-c1-multiasset | 32981093 | 2018-2025 | 0.258 | 6.490 | 17.0 | 8.7 | `1236cf59` | COMP framework C4.1 Multi-Asset Rotation (3-alpha ensemble MomentumAlpha+MACDAlpha+RelativeStrengthAlpha, RiskParityPCM equal-weight weekly 100% exposure 40% sector-cap, DrawdownCap 12%+trailing 4%, VWAP 4-slice execution; SPY/TLT/GLD/USO/EFA lev 2.0, IBKR margin). Sharpe 0.258 weak positive = 3-alpha ensemble adds little over single momentum (cf AssetClassMomentum 0.22, Vol-Ensemble 0.265); MaxDD 17% controlled by 12% drawdown cap. Far below framework leaders FamaFrenchAllWeather 0.684/EMATrend 0.741 = COMP composites vary widely, alpha/PCM choice > framework wiring. Aligned 0.258 > catalog 0.175 (alignment did not degrade). PSR 8.7% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| composite-c2-equityfactor | 32981222 | 2018-2025 | 0.574 | 11.942 | 18.6 | 25.765 | `8eecba32` | COMP framework C4.2 Equity Factor (4-alpha ensemble Value/Quality/LowVol/MomentumFactor, MeanVariancePCM weekly 65% exposure 18% sector-cap, SectorCapRiskModel 10% sector + 0.8 beta + trailing 4%, TWAP 6-slice execution; FineFundamentalUniverseSelectionModel coarse top-200 by dollar-volume then fine top-25 by market cap, IBKR margin). Sharpe 0.574 = STRONGEST COMP backbone verified, near Tier-1 (>0.5); CAGR 11.9% = 2nd-highest no-ML (after HAR-RV-J 14.1% #29). Holds & slightly improves vs catalog (0.543→0.574, PSR 16.8→25.8%) = robust not period-overfit. Contrasts sharply with sibling c1-multiasset 0.258 (#33): same Alpha Framework scaffold, but fine-fundamental top-25 stock factor investing dramatically outperforms static 5-ETF multi-asset rotation = component alpha/PCM choice + stock-level universe >> framework wiring. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| Framework_Composite_FamaFrenchAllWeather | 28882145 | 2018-2025 | 0.338 | 6.578 | 13.1 | 22.9 | `e9ac7c66` | COMP framework FamaFrench+AllWeather (FF20/AW80; FamaFrenchAlpha risk-adj momentum top-2 of VLUE/MTUM/SIZE/QUAL/USMV quarterly NO SMA200 + AllWeatherAlpha static SPY/IEF/GLD/XLP monthly, MultiStrategyPCM 20/80 monthly, NullRisk + ImmediateExecution, IBKR margin). Sharpe 0.338 = PERIOD-OVERFIT COLLAPSE: headline 0.588 (2010-2026 sweep) + OOS 0.684/PSR 87.5% (bt `b08c8956`, 835 dates) do NOT survive alignment — 0.684→0.338 (-51%), PSR 87.5→22.9% = small-OOS-sample artifact not robust edge; catalog 2015-2025 (bt `70415edc`) 0.472. The 80% AllWeather sleeve drags composite BELOW standalone FamaFrench 0.445 (#72) on aligned window BUT buys tight MaxDD 13.1% (among tighter backbones). Corrects the "FamaFrenchAllWeather 0.684 leader" refs in findings #33/#34 (that 0.684 was OOS-only). Promoted Tier 4→2. totalOrders=0 = wrapper artifact |
| Framework_Composite_EMATrend | 28911253 | 2018-2025 | 0.611 | 16.670 | 27.9 | 19.783 | `3095a263d5bd30df181ec002c0a52b72` | COMP framework EMA-Cross+TrendStocks (EMA70/Trend30 sweep WINNER; EMACrossAlpha 20/50 EMA on 5 Mag7 AAPL/MSFT/GOOGL/AMZN/NVDA + TrendStocksAlpha double-confirmation Price>SMA200+EMA20>EMA50 on 15 mega-caps incl the 5 Mag7, MultiStrategyPCM weekly, NullRisk + ImmediateExecution, IBKR margin). Sharpe 0.611 = SURVIVES alignment with a mild drop (catalog 0.741→0.611, -18%), NOT a period-overfit collapse (contrast #90 FamaFrenchAllWeather 0.588→0.338). HIGHEST-SHARPE COMP backbone on the aligned window (edges c2 0.574 #34, FamaFrenchAllWeather 0.338 #90, c1 0.258 #33); CAGR 16.67% = highest no-ML backbone. MAG7 SURVIVORSHIP CAVEAT: EMA sleeve 100% Mag7 → Sharpe partly a Mag7-decade artifact (loses pre-2018 ramp + absorbs 2022 drawdown on alignment); highest Sharpe ≠ most robust constitution — c2 0.574 (factor-diversified across 25 stocks) is the more defensible COMP leader constitution-wise, EMATrend is the higher-Sharpe Mag7-concentrated one. G.1: repo main.py had drifted to EMA40/Trend60 "starting point"; converged to the 70/30 deployed winner (matches QC Cloud 28911253 + catalog lines 36/119) for an apples-to-apples aligned run. Promoted Tier 4→2. totalOrders=0 = wrapper artifact |

| Project | QC ID | Period | Sharpe | CAGR% | MaxDD% | PSR% | Backtest ID | Notes |
|---------|-------|--------|--------|-------|--------|------|-------------|-------|
| RiskParity (inverse-vol aligned) | 33286158 | 2018-2024 | 0.361 | 8.304 | 18.4 | 11.3 | `0eaf6f5b` | RISK true inverse-volatility risk-parity (#46 catalog; rank-based RP v4 docstring "BEST": SMA200 trend filter + 12m-momentum rank × close/std60 inv-vol weighting, BND safe-haven <2 trending, monthly; SPY/EFA/GLD/DBC/TLT+BND, IBKR margin). Aligned 0.361 = -30% vs student 2015-2025 0.514 (proj 31872286 *non-aligned) = regime-4 fee-friction (bonds+commodities NOT fee-homogeneous, AllWeather -30% pattern). Catalog 0.40→0.361. Best risk-parity-family member (cf Cloud-RiskParity-Composite 0.027 #23 mis-named equal-weight). PSR 11.3% non-significant. Promoted Tier 4→2. totalOrders=0 = wrapper artifact (CAGR 8.3% ⇒ real trades) |
| DualMomentum (Antonacci aligned) | 33286487 | 2018-2024 | 0.350 | 8.413 | 14.9 | 9.6 | `b1cb13c9` | IND Antonacci canonical dual momentum (#48 catalog; absolute filter price>SMA200 AND 6m>0 + relative 12m ranking all-candidates momentum-tilt, BND safe-haven <2 trending, monthly; SPY/EFA/EEM/TLT/GLD/DBC+BND, IBKR margin). Aligned 0.350 = catalog 0.35 FLAT (no fee-collapse); tighter MaxDD 14.9% than sibling RiskParity 18.4% (dual-filter + BND refuge). PSR 9.6% non-significant — student 0.493/PSR 54.9 (proj 31798582, 2yr) was a small-sample artifact (finding #35 FamaFrenchAllWeather pattern). Promoted Tier 4→2. totalOrders=0 = wrapper artifact (CAGR 8.4% ⇒ real trades) |

### Student strategies (ESGF 5BD1 cohort, See #1405)

| Project | QC ID | Period | Sharpe | CAGR% | MaxDD% | PSR% | Backtest ID | Notes |
|---------|-------|--------|--------|-------|--------|------|-------------|-------|
| DualMomentum (student) | 31798582 | 2023-2025* | 0.493 | 13.5 | 9.0 | **54.9** | `88d36544` | *Hardcoded 2023 start. PSR > 50%! |
| RiskParity inverse-vol (student) | 31872286 | 2015-2025* | 0.514 | 9.3 | 20.7 | 16.3 | `d6a7bc52` | *Hardcoded 2015 start |
| ValueFactor Z-Score (student) | 31932810 | 2015-2025* | 0.227 | 6.4 | 36.5 | 0.8 | `da42c569` | Alpha negatif (decennie growth) |
| OptionWheel VGT (student) | 31846074 | 2018-2025* | -0.51 | 0% | 103.5 | 0.0 | `b9eca3c8` | Win-rate paradoxe, MaxDD > 100% |

**Note**: AdaptiveAssetAllocation (31781187) et MarkovRegime (31871247) n'ont produit aucune métrique (0 trades ou erreur d'exécution).

### Risk Parity pedagogique : inverse-vol vs ERC, et lecture critique du PSR (See #1405)

Les 4 strategies etudiantes ci-dessus sont aussi un support de cours. Cette section formalise les 3 points d'enseignement demandes par #1405 : (1) pourquoi le Sharpe seul ne suffit pas et ce que mesure le **PSR**, (2) la difference entre **inverse-volatilite naive** et **Equal Risk Contribution (ERC)**, (3) la lecture critique d'un backtest (dates hardcoded, MaxDD > 100%, fenetres non comparables).

#### 1. Le PSR (Probabilistic Sharpe Ratio) — Bailey & Lopez de Prado (2012)

Le ratio de Sharpe observe suppose des rendements **IID et gaussiens**. Les rendements réels violent ces deux hypotheses : ils ont en general une **asymetrie (skew) negative** (les krachs sont plus profonds que les booms ne sont hauts) et des **queues epaisses (kurtosis > 3)**. Sous ces conditions, le Sharpe empirique est **systematiquement surestime**, surtout sur de courtes fenetres.

Le PSR corrige en estimant la probabilité que le vrai Sharpe depasse un seuil de reference `SR0` (souvent 0, le hasard) :

```text
                     (SR_hat - SR0) * sqrt(T - 1)
   PSR(SR0) = Phi( ------------------------------------ )
                  sqrt( 1 - skew*SR_hat + (kurt-1)/4 * SR_hat^2 )
```

ou `T` = nombre d'annees d'observation (quand `SR_hat` est annualise), `skew` et `kurt` sont les moments empiriques des rendements, et `Phi` = fonction de repartition de la loi normale centree.

**Lecture** : asymetrie negative et exces de kurtosis (kurt > 3) **gonflent le denominateur**, donc abaissent le PSR. Un meme Sharpe nominal peut donner un PSR tres different selon la distribution des rendements. Regle pratique : **PSR > 50%** = l'edge observe a plus de chances d'etre réel que d'etre du bruit ; **PSR < 50%** = on ne peut pas ecarter le bruit.

**Applique aux 4 strategies etudiantes** :

| Strategie | Sharpe | PSR% | Lecture |
|-----------|--------|------|---------|
| DualMomentum | 0.493 | **54.9** | Edge a la limite de la significativite (juste au-dessus de 50%). Fenetre courte (2 ans) => estimation des moments bruitee, a confirmer sur une periode plus longue. |
| RiskParity inverse-vol | 0.514 | 16.3 | Sharpe plus haut mais PSR faible : sur 10 ans, l'edge moyen est réel mais **statistiquement peu concluant** (estimation robuste d'un edge faible). |
| ValueFactor | 0.227 | 0.8 | Quasi-zero : l'alpha observe est indistinguable du bruit. Confirme l'effondrement du facteur value sur une decennie dominee par la growth. |
| OptionWheel | -0.51 | 0.0 | Aucun edge. La probabilité d'un vrai Sharpe positif est nulle. |

**Contraste cle** : Sharpe et PSR ne classent pas pareil. RiskParity bat DualMomentum en Sharpe (0.514 > 0.493) mais DualMomentum le bat largement en PSR (54.9 > 16.3). Le PSR est la statistique a citer, pas le Sharpe brut — premiere lecon de lecture critique.

#### 2. inverse-volatilite naive vs Equal Risk Contribution (ERC)

Le `RiskParity inverse-vol` etudiant n'implemente pas du "vrai" risk parity. Deux familles distinctes :

**inverse-volatilite (naive)** — ce que fait l'etudiant :

```text
   poids_i  proportionnel a  1 / sigma_i
```
Chaque actif recoit un poids inverse a sa volatilite **individuelle**. Methode simple, une seule donnée par actif, mais elle **ignore les correlations**.

**Equal Risk Contribution (ERC)** — Maillard, Roncalli & Teiletche (2010), le "vrai" risk parity :

```text
   chaque actif i contribue EGALEMENT au risque total :
   (Sigma * w)_i / (w' * Sigma * w)  =  constant   pour tout i
```
ou `Sigma` est la **matrice de covariance** complete (correlations incluses). Résolution par programmation convexe (QP / Newton-Lagrange), sans solution analytique en general.

**Pourquoi la difference compte** : avec l'inverse-vol naive, deux actifs **fortement correles** (ex. deux ETF actions US) reçoivent chacun un budget de risque eleve, donc le portefeuille reste **concentre sur un meme facteur** — faussement "diversifie". L'ERC, en integrant la covariance, **force une vraie diversification** : deux actifs correles se partagent un meme budget de risque cumule, pas deux budgets independants.

**Pedagogique** : le `RiskParity inverse-vol` (Sharpe 0.514, PSR 16.3%) est un **bon point d'entrée** — simple, robuste, peu de paramètres. L'upgrade naturel est l'**ERC** (gestion des correlations via la matrice de covariance). Le saut conceptuel : passer de "donner moins de poids au plus volatile" (1D) a "egaliser la contribution marginale au risque" (matricielle, multidimensionnelle).

#### 3. Lecture critique d'un backtest — les pieges visibles dans la table ci-dessus

- **Dates de debut hardcoded** (`*` sur les 4) : les fenetres ne sont **pas aligned** avec le reste du catalogue (2018-2025). Les metriques ne sont pas directement comparables — comparer le Sharpe d'un DualMomentum 2023-2025 a un TrendFollowing 2018-2025 est abusif.
- **MaxDD > 100% (OptionWheel, 103.5%)** : un drawdown superieur a 100% signale une vente d'options **naked non couverte** integrale — le simulateur ne capture pas parfaitement l'assignation / liquidation forcee. Le backtest est probablement **optimiste** sur la perte réelle.
- **Croiser PSR et duree** : un PSR de 54.9% sur 2 ans (DualMomentum) n'a pas le meme poids qu'un PSR de 81.8% sur 8 ans (TrendFollowing, ligne 381). La signification statistique croit avec `sqrt(T)` ; une courte fenetre a besoin d'un edge plus fort pour convaincre.

#### References

- Bailey, D. & Lopez de Prado, M. (2012), *"The Sharpe Ratio Efficient Frontier"*, Journal of Risk.
- Lopez de Prado, M. (2014), *"The Deflated Sharpe Ratio"*, SSRN (extension corrigeant le biais de selection multiple).
- Maillard, S., Roncalli, T. & Teiletche, J. (2010), *"The Properties of Equally Weighted Risk Contribution Portfolios"*, Journal of Portfolio Management.
- Broad, J., *Hands-On AI Trading* (chap. performance / risk-adjusted metrics).

### Not alignable (hardcoded ML train/test split)

| Project | QC ID | Reason | Current Period |
|---------|-------|--------|----------------|
| BTC-ML | 29318876 | Train 2019-2022, test 2023-2026 hardcoded. Changing dates breaks ML logic. | 2023-01-01 → 2026-03-01 |

---

## Key findings

### Verdicts initiaux de la campagne alignée 2018-2025 (#1-#9)

1. **TrendFollowing = leader indiscutable**: Sharpe 1.072 sur 2018-2025 avec MaxDD 9.3%. PSR 81.8% (statistiquement significatif). Seule strategie "Robuste" confirmee sur la periode aligned.

2. **EMA-Cross-Stocks: surprise positive**: Sharpe 0.891 sur 2018-2025, CAGR 26.2%. PSR 40.5%. 2e meilleur Sharpe aligne, derriere TrendFollowing.

3. **EMA-Cross-Alpha: chute dramatique**: Sharpe passe de 0.996 (meilleur backtest) a -0.010 sur la periode aligned. PSR 0.5% = bruit. Confirme le pattern "backtests courts = overfitting".

4. **Composites ne battent pas les single-strategies**: MomentumRegime (SectorMomentum + RegimeSwitching) obtient 0.185, confirmant le probleme de "double-defense".

5. **Crypto = rendement modere mais stable**: Crypto-MultiCanal (0.581) et Portfolio-IBKR-Binance (0.519) offrent diversification avec MaxDD maitrises.

6. **FX Carry = perdant**: Sharpe -1.108, les taux bas post-COVID ont elimine l'avantage du carry trade.

7. **AllWeather: performance confirme**: Sharpe 0.631 (2010-2025), MaxDD 16.4%, PSR 31.2%. Risk-parity solide.

8. **MomentumStrategy (SectorMom v4.0)**: Sharpe 0.555, CAGR 11.7%, mais PSR 7.2% = non significatif.

9. **SectorDualMomentum v3.2**: Sharpe 0.581, CAGR 13.5%, MaxDD 22.8%. PSR 15.3%.

### Catalogue re-vérifié et stratégies étudiantes (#10-#21)

10. **TrendStocks-Alpha: high return, high risk**: CAGR 15.9% mais MaxDD 39.6% (Calmar 0.40). PSR 5.8% = non significatif.

11. **Student DualMomentum: PSR > 50%**: Sharpe 0.493 sur 2023-2025 avec MaxDD 9.0%. Seule strategie etudiante avec PSR significatif (54.9%).

12. **Student RiskParity: performance honnete**: Sharpe 0.514, CAGR 9.3%, MaxDD 20.7%. Inverse-vol simple mais efficace. PSR 16.3% (non significatif mais respectable).

13. **Student OptionWheel: catastrophe pedagogique**: Sharpe -0.51, MaxDD 103.5%. Parfait comme etude de cas du "win-rate paradoxe".

14. **Student ValueFactor: alpha negatif confirmee**: Sharpe 0.227, PSR 0.8%. Decennie growth-dominée = facteur value sous-performant.

15. **LeveragedETFMomentum: confirme et significatif**: Sharpe 1.779 sur 2018-2025, PSR 79.8% (2e PSR significatif apres TrendFollowing). Mais MaxDD 53.3% et CAGR 126% typiques d'un levier 3x — profil risque extreme, pas comparable aux strategies non-leveragees.

16. **PuppiesOfTheDow et HighBookToMarketFScore: effondrement sur periode alignee**: Sharpe catalog 1.99 et 2.09 (obtenus sur leur fenetre glissante par defaut `end_date - 12 ans`) tombent a 0.302 (PSR 3.5%) et 0.411 (PSR 4.5%, MaxDD 60.4%) sur 2018-2025. Les deux meilleures lignes ML/IND du Tier 1 ne sont pas reproductibles sur la fenetre standardisee.

17. **TrendWeather: le composite qui tient**: Sharpe 0.948 (PSR 56.6%), proche du catalog 1.16. Contraste fort avec MomentumRegime (0.185) — toutes les architectures composites ne se valent pas.

18. **Caveat reproductibilite Trend-Following**: le code du repo backteste sur 2018-2024 donne Sharpe 0.365 / MaxDD 13.8% (backtest `3748cb62`), loin du 1.072 publie ci-dessus (`7792ae0a`, 2018-2025, etat du code cloud anterieur). Periodes differentes (2025 inclus ou non) ET drift possible repo vs cloud — a investiguer avant de citer 1.072 comme reference du code versionne.

19. **MeanReversion v5.2: Best Calmar ratio**: Sharpe 0.81, MaxDD 7.5%, Calmar 1.34 — best risk-adjusted return among non-leveraged strategies. PSR 46.8% (near significance). Promoted from Tier 2 (0.29) to Tier 1. The v5.2 code (IBKR brokerage, RSI65 exit, 10% stop-loss) dramatically outperforms the older version.

20. **AdaptiveAssetAllocation: confirmed robuste**: Sharpe 0.509, CAGR 8.0%, MaxDD 18.9% (2008-2024, 16 years). Min-var + momentum approach produces steady returns. PSR 10.6% (not significant but positive).

21. **PairsTrading: structural failure confirmed**: Sharpe -0.28 on aligned period, PSR 0.001%. OLS hedge + cointegration still produces negative alpha. Remains exploratoire/pedagogical.

### Première vague de backbones no-ML, 2026-06-22 (#22-#28)

22. **AssetClassMomentum-QC: weak aligned baseline (2026-06-22)**: 5-ETF momentum (top-3 of SPY/EFA/BND/VNQ/GSG, 252d lookback, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.22, CAGR 6.6%, MaxDD 28.1%, PSR 3.8% (non-significant). Confirms the aligned-period momentum underperformance pattern (cf MomentumRegime 0.185, EMA-Cross-Alpha -0.010). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `6746f155`, project 33209767.

23. **Cloud-RiskParity-Composite: near-flat aligned baseline (2026-06-22)**: 6-asset tactical rotation (SPY/TLT/GLD/EFA/EEM/DBC, SMA200 + 6m momentum dual filter, equal weight, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.027, CAGR 3.5%, MaxDD 24.4%, PSR 1.2% (non-significant). Despite the "RiskParity" name the code is equal-weight momentum rotation (AQR Trend-Following style, not true risk-parity weighting). Extends the aligned-period rotation/momentum underperformance pattern (cf AssetClassMomentum 0.22 #22, Cloud-MeanReversion 0.067). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `b184b13e`, project 30820857. Note: `totalOrders=0` in the MCP wrapper is an extraction artifact (CAGR 3.5% ⇒ real trades), not a 0-trade backtest; cross-checked `list_backtests` status Completed.

24. **Cloud-SectorRotation-Momentum: first Tier-3 (negative) aligned baseline (2026-06-22)**: 5-ETF momentum-weighted rotation (QQQ/SPY/EFA/GLD/IWM, SMA200 + 6m momentum dual filter, momentum-proportional sizing, SHY defensive, monthly rebalance, IBKR) on 2018-2025 gives Sharpe -0.029, CAGR 2.1%, MaxDD 42.7%, PSR 0.5% (non-significant). The momentum-weighted variant does *slightly worse* than the equal-weight Cloud-RiskParity-Composite #79 (-0.029 vs +0.027) — momentum-proportional sizing did not help on the aligned period. Strengthens the aligned-period momentum-underperformance finding (first baseline below zero). Promoted Tier 4 (Untested) → Tier 3 (Exploratoire). Backtest `58fb0a94`, project 30821748. Same `totalOrders=0` wrapper extraction artifact as #77/#78/#79 (CAGR 2.1% ⇒ real trades).

25. **Cloud-VolTargeting: true vol-targeting, leverage-clamp raises MaxDD (2026-06-22)**: single-asset SPY vol-targeting (target_vol 12%, 21d realized-vol lookback, 30-150% allocation clamp, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.207, CAGR 6.7%, MaxDD 38.2%, PSR 2.4% (non-significant). Genuine vol-targeting (unlike the misnomer Cloud-RiskParity-Composite #79 which is equal-weight rotation). Notable: MaxDD 38% *exceeds* AssetClassMomentum #77 (28%) — the 150% upper clamp leveraged the portfolio into vol spikes (vol↑ ⇒ target/realized demands >100% ⇒ clamped at 150% ⇒ amplified drawdown). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `179ef1c5`, project 30823587. Same totalOrders=0 wrapper artifact.

26. **GlobalMacro-Regime: best backbone baseline, regime-switch + risk-parity wins (2026-06-22)**: rank-based risk-parity + SPY regime switch (Bridgewater / Antonacci 2014; SMA200+6m-mom bull/bear gate, rank×inv_vol weighting over trending risky assets in bull, inv-vol risk-parity over BND/TLT/GLD defensive in bear, monthly rebalance, IBKR) on 2018-2025 gives Sharpe **0.454**, CAGR 9.8%, MaxDD 22.8%, PSR 16.7% (respectable). The strongest aligned backbone baseline yet — nearly Tier 1 (>0.5) — and the regime-switch + inverse-vol risk-parity combination dramatically outperforms simple momentum rotations (AssetClassMomentum 0.22 #22, Cloud-RiskParity 0.027 #23, Cloud-SectorRotation -0.029 #24) while *also* keeping MaxDD controlled (22.8% vs #80 42.7%, #81 38.2%). The defensive regime rotation (risk-parity over BND/TLT/GLD in bear markets) is the likely edge over pure trend/momentum. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `bbb73b7b`, project 30781695. Same totalOrders=0 wrapper artifact.

27. **MomentumRegime-AdaptiveWeights: double-defense composite destroys the SectorMomentum edge (2026-06-22)**: COMP framework composite (SectorMomentum 85% + RegimeSwitching 15% via QC Alpha / PortfolioConstruction additive; composite momentum 1/3/6/12m weights 0.5/0.2/0.2/0.1 + SMA200 filter; RegimeSwitching SPY SMA50/SMA200 bull/bear/sideways + RSI; MultiStrategyPCM groups by source_model; SPY/QQQ/IEF/GLD universe, IBKR) on 2018-2025 gives Sharpe **-0.729**, CAGR 1.875%, MaxDD 4.3%, PSR 17.4% (non-significant). The composite was *overwhelmingly defensive* on the aligned period (SMA200 filter + bear/sideways regime → mostly IEF/GLD/cash), so MaxDD is tiny (4.3%) but CAGR 1.875% < risk-free → negative Sharpe. The T85/RS15 variant is *WORSE* than the baseline 60/40 MomentumRegime composite (Sharpe 0.185) — shifting weight toward SectorMomentum added defensiveness, not edge. Confirms + extends the double-defense finding (cf Key-finding #4): standalone SectorMomentum works (Tier 1, 0.56 #22), but wrapping it in a regime-gated composite destroys the edge on the aligned period. Promoted Tier 4 (Untested) → Tier 3 (Exploratoire) — 2nd Tier-3 negative. Backtest `6e8f164f`, project 31524424. Same totalOrders=0 wrapper artifact.

28. **TermStructureCommodities-QC: roll-yield signal catastrophically fails on the modern period (2026-06-22)**: long-short commodity futures on roll returns / backwardation-contango (top-quintile backwardation = long, top-quintile contango = short; near vs distant contract log-price ratio annualized by expiry gap; monthly rebalance; 21 commodity futures across Softs / Grains / Meats / Energies / Metals; IBKR margin) on 2018-2025 gives Sharpe **-0.244**, CAGR **-31.5%**, MaxDD **96.8%**, PSR 0.007% (non-significant) — the worst backbone baseline yet, nearly going to zero. This *confirms the published library header* (OOS 5Y Sharpe -0.041, MaxDD 80.8%): the roll-yield / backwardation signal that worked historically catastrophically fails on the modern period. The 2020 COVID oil crash + 2022 energy / inflation spike dislocated term structures (extreme contango then extreme backwardation), putting the strategy on the wrong long-short legs through both regimes — and the cherry-picked "Recent OOS 1.49 1Y Sharpe" sub-period in the header does not survive alignment. Reinforces the #1630 finding: published headline Sharpe figures do not survive alignment to a modern, fee-aware period. Promoted Tier 4 (Untested) → Tier 3 (Exploratoire) — 3rd Tier-3 negative (after #80 -0.029, #86 -0.729), worst by CAGR / MaxDD. Backtest `e9f7e686`, project 33224097 (created fresh; original library project 29688398 absent). Same totalOrders=0 wrapper artifact.

### Familles RISK/COMP/FACTOR — 2026-06-22/23 (#29-#33)

29. **HAR-RV-J-Kelly: crypto vol-forecasting + Kelly survives alignment (2026-06-22)**: HAR-RV-J volatility forecasting (Corsi 2009 + Andersen-Bollerslev-Diebold 2007 jump component via Huang-Tauchen bipower variation; inline OLS via np.linalg.lstsq, refit every 22 days; iterated 5-day-ahead log-RV forecast; 1/4-Kelly position sizing × 5-day-momentum direction; BTC/ETH/LTC/BCH USDT on Binance) on 2018-2025 gives Sharpe **0.524**, CAGR 14.1%, MaxDD 37.1%, PSR 10.7% (non-significant). The **strongest backbone baseline since GlobalMacro-Regime (0.454, #26)**, and notably the first RISK-family baseline whose signal survives alignment with a *positive* Sharpe — contrasting sharply with TermStructureCommodities (#28, Sharpe -0.244 / MaxDD 96.8%): where the commodity roll-yield signal catastrophically broke on the modern period, the crypto jump-aware HAR-RV-J vol model + fractional Kelly held. Counter-intuitive MaxDD note: extending 2020-2025 → 2018-2025 *lowered* MaxDD (48.3% → 37.1%) — the longer OLS coefficient path through the 2022 crypto bear produced a less-leveraged exposure than the shorter window. PSR 10.7% = non-significant (crypto single-directional long-only + 1/4-Kelly cap keeps it conservative). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `df687834`, project 31650567. Same totalOrders=0 wrapper extraction artifact (CAGR 14.1% ⇒ real trades).

30. **Vol-GARCH-Target: GARCH vol-targeting delivers the tightest drawdown of the backbone (2026-06-23)**: GARCH(1,1) volatility targeting on 6 multi-asset ETFs (SPY/EFA/EEM/TLT/GLD/DBC; variance-targeting MLE via inline alpha/beta grid search maximizing the Gaussian log-likelihood, pure numpy, no `arch` dependency; refit every 22 days on a 500-day window; 10% annualized vol target per asset slot, 30% per-asset cap, SMA200 trend direction, weekly Monday rebalance, IBKR) on 2018-2025 gives Sharpe **0.325**, CAGR 6.97%, MaxDD **10.80%**, PSR 14.9% (non-significant). The defining feature is **risk control, not return**: MaxDD 10.80% is the **tightest of all backbone baselines** to date (cf GlobalMacro-Regime 22.8% #26, HAR-RV-J-Kelly 37.1% #29, Cloud-VolTargeting single-asset 38.2% #25) — the GARCH variance forecast, 30% per-asset cap and SMA200 trend filter combine into genuine risk budgeting. It also **beats single-asset Cloud-VolTargeting #25** (Sharpe 0.325 vs 0.207, MaxDD 10.8% vs 38.2%): GARCH forecasting plus multi-asset diversification outperforms a naive 21d realized-vol single-asset with a 150% leverage clamp — confirming the value of a proper variance model over a leverage-clamped heuristic. PSR 14.9% non-significant (conservative vol-capped long-only). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `a0f7a5b5`, project 33245149 (created fresh; no prior library project). Same totalOrders=0 wrapper extraction artifact (CAGR 6.97% ⇒ real trades).

31. **Vol-Ensemble-Conservative: the conservative ensemble tightens drawdown further but HAR adds no Sharpe over GARCH alone (2026-06-23)**: conservative ensemble volatility forecasting (max of GARCH(1,1) + HAR(1,5,22) forecasts; GARCH variance-targeting MLE inline alpha/beta grid + HAR OLS np.linalg.lstsq, both pure numpy, no `arch`; 8% annualized vol target half-allocation, 25% per-asset cap, SMA200 regime -50% in bear, SMA50 direction, weekly Monday rebalance; 6 ETFs SPY/EFA/EEM/TLT/GLD/DBC, IBKR) on 2018-2025 gives Sharpe **0.265**, CAGR 6.142%, MaxDD **10.40%**, PSR 13.6% (non-significant). MaxDD 10.40% is the **new tightest of all backbone baselines** (beats Vol-GARCH-Target 10.80% #30, GlobalMacro-Regime 22.8% #26, HAR-RV-J-Kelly 37.1% #29). But the Sharpe (0.265) is **below single-GARCH Vol-GARCH-Target (0.325)**: the added conservative overlay (max-of-ensemble forecast, half vol-target 8%, SMA200 regime filter -50% in bear, SMA50 direction) tightened the drawdown by 40 bps at the cost of return, and the HAR component added **no risk-adjusted value over GARCH alone** on this period — a clean counter-example to "more sophisticated vol model = better risk-adjusted return" within the RISK family. PSR 13.6% non-significant. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `db77ae49`, project 33248352 (created fresh; the repo's stale cloud-id 31456204 points to the unrelated ESGF student project ESGF-VolEnsemble-Conservative). Same totalOrders=0 wrapper extraction artifact (CAGR 6.142% ⇒ real trades).

32. **FamaFrench: risk-adjusted factor momentum survives alignment, robust not period-overfit (2026-06-23)**: risk-adjusted momentum ETF rotation (FactorETFRotation v3.0; 12m-1m return divided by 63d realized volatility = momentum-Sharpe score, Barroso & Santa-Clara 2015; dynamic top_n = all factors with positive score; SMA200 SPY regime filter, USMV risk-off in bear, per-position -12% stop-loss, monthly rebalance; 5 Fama-French factor ETFs VLUE/MTUM/SIZE/QUAL/USMV, IBKR) on 2018-2025 gives Sharpe **0.445**, CAGR 11.111%, MaxDD 24.1%, PSR 11.9% (non-significant). The 3rd-best no-ML backbone (close to GlobalMacro-Regime 0.454 #26) with the 2nd-highest CAGR (11.1%, after HAR-RV-J-Kelly 14.1% #29). The strategy **survives alignment with only a mild Sharpe drop vs the author's 2015-2024 v3.0** (0.540 → 0.445; MaxDD stable 24.2% → 24.1%) — genuinely robust, not period-overfit, in **contrast to the broader "simple factor collapse" finding**: this is a risk-adjusted momentum rotation (momentum-Sharpe ranking + regime filter + stop-loss), not static factor exposure, and that risk-adjustment is what keeps it positive on the modern period. It sits **below the FamaFrenchAllWeather composite (Sharpe 0.684)**, confirming the framework composite adds value over the standalone rotation. PSR 11.9% non-significant. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `697e96af`, project 33251801 (created fresh; repo had no cloud project). Same totalOrders=0 wrapper extraction artifact (CAGR 11.1% ⇒ real trades).

33. **composite-c1-multiasset: COMP framework ensemble adds little over single momentum, alpha/PCM choice > wiring (2026-06-23)**: QC Alpha Framework C4.1 Multi-Asset Rotation composite (3-model alpha ensemble: MomentumAlpha 12m-1m risk-adjusted + MACDAlpha crossover + RelativeStrengthAlpha cross-asset 3m; RiskParityPCM equal-weight weekly 100% max-exposure 40% sector-cap; DrawdownCap 12% + trailing-stop 4%; VWAP 4-slice execution; universe SPY/TLT/GLD/USO/EFA at 2x leverage, IBKR margin) on 2018-2025 gives Sharpe **0.258**, CAGR 6.490%, MaxDD 17.0%, PSR 8.7% (non-significant). A weak positive Sharpe — the 3-alpha ensemble + framework architecture adds **little over a single momentum signal** on the aligned period (cf AssetClassMomentum 0.22 #22, Vol-Ensemble-Conservative 0.265 #31). MaxDD is controlled at 17% by the 12% drawdown cap (vs 22.8% for the uncapped GlobalMacro-Regime #26). It sits **far below the framework composite leaders** FamaFrenchAllWeather (0.684) and EMATrend (0.741): **COMP composites vary widely, so the component alpha/PCM choice matters more than the framework wiring itself** — the same Alpha Framework scaffold produces a 0.258 weakling here vs 0.684-0.741 leaders depending on which alphas and PCM are plugged in. Notably the aligned Sharpe 0.258 is *higher* than the catalog-campaign figure (0.175), so alignment did not degrade this composite. PSR 8.7% non-significant. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `1236cf59`, project 32981093 (pre-existing; 5 framework files verified present, only main.py date-aligned). Same totalOrders=0 wrapper extraction artifact (CAGR 6.49% ⇒ real trades).

### Analyses COMP approfondies et vetting ML — 2026-06-23 (#34-#40)

**34. composite-c2-equityfactor: strongest COMP backbone — fine-fundamental stock factor investing dominates static ETF rotation (2026-06-23)**

QC Alpha Framework C4.2 Equity Factor composite (4-model alpha ensemble: ValueAlpha P/E+P/B, QualityAlpha ROE+debt-to-equity, LowVolAlpha realized-vol, MomentumFactorAlpha price momentum; MeanVariancePCM weekly 65% max-exposure 18% sector-cap; SectorCapRiskModel 10% sector weight + 0.8 beta + trailing 4% stop; TWAP 6-slice execution; FineFundamentalUniverseSelectionModel coarse top-200 by dollar-volume then fine top-25 by market cap, IBKR margin) on 2018-2025 gives Sharpe **0.574**, CAGR 11.942%, MaxDD 18.6%, PSR 25.8% (non-significant but the highest PSR among the no-ML backbones). The **strongest COMP (composite-framework) backbone verified to date** — near Tier-1 (>0.5) — and it **holds and slightly improves** over the catalog window (0.543 → 0.574, PSR 16.8% → 25.8%), genuinely robust rather than period-overfit.

The defining contrast is with the sibling **composite-c1-multiasset (Sharpe 0.258, #33)**: both share the *same* Alpha Framework scaffold, yet c2 more than doubles c1's Sharpe. The difference is the **universe and factors, not the wiring**: c2 invests in individual large-cap US stocks selected by fine fundamentals and scored across Value/Quality/LowVol/Momentum (true factor investing across 25 single-name equities with a mean-variance PCM), whereas c1 rotates 5 static asset-class ETFs on a momentum/MACD/relative-strength ensemble. Within the COMP family this confirms and sharpens the #33 finding — **the component alpha/PCM choice and the stock-level fundamental universe matter far more than the framework wiring itself**, and a fine-fundamental stock universe is where the Alpha Framework actually earns its Sharpe.

It remains below the framework leaders FamaFrenchAllWeather (0.684) and EMATrend (0.741), but closes most of the gap c1 left open.

Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `8eecba32`, project 32981222 (pre-existing catalog project `c2-equityfactor-post2801` bt `3173f8b39`; only main.py date-aligned). Same totalOrders=0 wrapper extraction artifact (CAGR 11.9% ⇒ real trades).

**35. FamaFrenchAllWeather: headline collapses on alignment — the 0.684 / PSR 87.5% OOS figure was a small-sample artifact (2026-06-23)**

QC Alpha Framework composite FamaFrench (20%) + AllWeather (80%) (FamaFrenchAlpha risk-adjusted-momentum top-2 rotation over VLUE/MTUM/SIZE/QUAL/USMV, skip-month, quarterly, NO SMA200 filter — AllWeather handles defense; AllWeatherAlpha static SPY/IEF/GLD/XLP Ray-Dalio-inspired monthly drift-rebalance; MultiStrategyPCM 20/80 allocation monthly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.338**, CAGR 6.578%, MaxDD 13.1%, PSR 22.9% (non-significant).

**A clear period-overfitting collapse**: the headline figures do *not* survive alignment. The README's 2010-2026 sweep Sharpe 0.588 falls to 0.338 (-42%), and — most tellingly — a separately-run OOS 2023-2026 backtest (`b08c8956`) that reported Sharpe 0.684 / PSR **87.5%** on only **835 tradeable dates** collapses to 0.338 / PSR 22.9% on the full 1761-date aligned window: the 87.5% PSR was a **small-OOS-sample artifact**, not a robust statistical edge. The same-config catalog FF20/AW80 over 2015-2025 (`70415edc`) = 0.472.

On the aligned window the 80% AllWeather sleeve (mostly static SPY/IEF/GLD/XLP) drags the composite *below* the standalone FamaFrench rotation (0.445, #32) — though it buys tight drawdown control (MaxDD 13.1%, among the tighter backbones alongside Vol-GARCH 10.8% #30 / Vol-Ensemble 10.4% #31).

**This corrects the "FamaFrenchAllWeather 0.684 framework leader" references in findings #33 and #34** — that 0.684 was the short OOS window, not the aligned baseline; the real aligned COMP leaderboard is composite-c2-equityfactor 0.574 (#34), not FamaFrenchAllWeather.

Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `e9ac7c66`, project 28882145 (pre-existing; cloud main.py was an OOS 2023-2026 variant, replaced with the canonical repo version date-aligned to 2018-2025). Same totalOrders=0 wrapper extraction artifact (CAGR 6.6% ⇒ real trades).

**36. EMATrend: survives alignment (0.741 → 0.611) — the highest-Sharpe COMP backbone, but Mag7-concentrated (survivorship caveat) (2026-06-23)**

QC Alpha Framework composite EMA-Cross (70%) + TrendStocks (30%) (EMACrossAlpha 20/50 EMA on the 5 Mag7 AAPL/MSFT/GOOGL/AMZN/NVDA, daily emission; TrendStocksAlpha double-confirmation Price>SMA200 + EMA20>EMA50 on 15 mega-caps including the same 5 Mag7, weekly; MultiStrategyPCM 70/30 weekly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.611**, CAGR 16.670%, MaxDD 27.9%, PSR 19.8% (non-significant).

**Survives alignment with a mild drop** (catalog 0.741 @ 2015-2025 → 0.611, -18%) — NOT a period-overfit collapse, in direct contrast to the sibling framework composite FamaFrenchAllWeather (#35, 0.588→0.338). The drop is the expected cost of losing the 2015-2017 Mag7 pre-ramp and absorbing the 2022 Mag7 drawdown; the trend signal itself holds.

On the aligned window this is the **highest-Sharpe COMP backbone verified to date**, edging composite-c2-equityfactor 0.574 (#34), and it carries the highest CAGR (16.67%) of any no-ML backbone.

**But the Mag7 survivorship caveat dominates the interpretation**: the EMA sleeve is 100% Mag7, so a meaningful fraction of the Sharpe is an artifact of the Mag7 outperformance regime that defined the 2015-2025 decade rather than a transferable trend-following edge. Highest Sharpe is **not** the same as the most robust constitution — composite-c2-equityfactor (0.574, fine-fundamental factor investing across 25 large-cap stocks) is the more defensible COMP leader constitution-wise; EMATrend is the higher-Sharpe but Mag7-concentrated one.

G.1 note: the repo `main.py` had drifted to the EMA40/Trend60 sweep *starting point* (its own docstring), while the deployed QC Cloud project (28911253) and the catalog entry both use the 70/30 sweep *winner* — the repo was converged to the winner so the aligned run is apples-to-apples vs the catalog 0.741.

Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `3095a263d5bd30df181ec002c0a52b72`, project 28911253. Same totalOrders=0 wrapper extraction artifact (CAGR 16.7% ⇒ real trades).

37. **Research-Executor is a research harness (not a strategy) — the #1630 no-ML backbone campaign is complete (2026-06-23)**: a G.1 verification of the queued "#1630 next floor = #92 Research-Executor (dernier leader COMP à vérifier)" against the source reveals Research-Executor is a **research execution harness, not a tradable strategy**. Its `main.py` sets a 2-day window (2024-01-02 → 2024-01-03), runs 8 embedded research notebooks via a `MockQB` shim, writes the executed notebooks to the object store, and calls `self.quit('Done')` inside `initialize` (with `on_data = pass`) — there is no portfolio, no PnL and no tradable mechanics, so an "aligned 2018-2025 baseline" is meaningless (the algorithm quits on day 1 regardless of the date range). Its README and `projects/catalog.json` already classify it `Type: Utility (research execution harness, not a trading strategy)` / `classification: untested`; it was **miscategorized** in the deep-queue as a COMP leader. This **closes the #1630 no-ML backbone campaign**: every IND/COMP/RISK/FACTOR candidate (#72, #77–#91) is verified across Key-findings #22–#36 (15 baselines), with the aligned-COMP leaderboard led by EMATrend 0.611 (#36, highest Sharpe, Mag7 survivorship caveat) and composite-c2-equityfactor 0.574 (#34, the most robust constitution). The remaining Tier-4 entries are characterized in the table note above (OPT naked-options with a MaxDD > 100 % simulator caveat; ML/DL/RL deferred to multi-seed #6).

**38. Adaptive-Conformal-Risk: real ACI overlay survives real fees (low-turnover immunity), but the Sharpe is Mag7 survivorship, not conformal value (2026-06-23)**

Adaptive Conformal Inference risk overlay on multi-factor momentum (ACI algorithm, Gibbs & Candès 2021: online alpha-adjustment `alpha_{t+1} = alpha_t + gamma·(1{violation} − target_alpha)` over a 60-day rolling nonconformity-score window, empirical-quantile prediction interval with finite-sample `1+1/√n` correction; position sizing `signal·confidence/interval_width` — wider conformal interval → smaller position; 3-window momentum signal 21/63/126d, sector cap 30%, target-vol 15%, monthly rebalance; 15 large-caps across 5 sectors, IBKR margin; source: ECE student project El Bakkali Gr02, Issue #238) on 2018-2025 gives Sharpe **0.449**, CAGR 11.525%, MaxDD 22.5%, PSR 12.526% (non-significant), 2011 tradeable dates.

This is the **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: the catalog 0.42 was unverified (it matches no Cloud backtest), and the only pre-existing Cloud run ("ACI-Risk-v1", project 29841071, Sharpe 0.604 / 2015-2026) had **silently lost its `set_brokerage_model(IBKR)` line** in the deployed copy and ran on the negligible default fee model — i.e. not #1630-compliant.

Two findings, one positive and one a caveat:

**(a) the low-turnover fee-resistance pattern holds** — the aligned-with-IBKR 0.449 sits only modestly below the no-fee 0.604, and much of that gap is the harder 2018-2025 window (no 2015-2017 bull-momentum tailwind) rather than pure fee drag; the ACI vol-targeting actively *scales exposure down*, cutting turnover, which *helps* fee-resistance. A genuinely sophisticated conformal-prediction risk overlay does **not** collapse under real fees — reinforcing the #1630 realized-turnover discriminator (cf SectorMomentum 0.56 #22, BlackLitterman 0.83 near-immune): it is turnover that kills strategies, not mechanism sophistication.

**(b) But the Mag7-survivorship caveat dominates the interpretation** — the universe is AAPL/MSFT/NVDA/AMZN/TSLA (5 of 7 Mag7) plus 10 financials/healthcare/consumer/industrial large-caps, and over 2018-2025 the Mag7 (NVDA especially) ran enormously, so long-only momentum concentrates in them and the 0.449 is substantially survivorship drift rather than ACI-overlay marginal value (the ACI vol-targeting may even *drag* vs uncapped Mag7 momentum). The overlay's contribution cannot be isolated without a uniform-sizing control — same caveat class as EMATrend (#36, Mag7 sleeve).

Net: a real, fee-resistant SOTA mechanism deployed on a survivorship-biased universe → the honest headline is "0.449, but it is mostly Mag7, not conformal prediction."

Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified). Backtest `cce5af0271e5e5989908897b6c5ecb08`, project 33278416 (baseline-clone `1630-baseline-AdaptiveConformalRisk`, created fresh; only main.py date-aligned to 2018-2025 + IBKR brokerage restored to match the canonical local copy). Same totalOrders=0 wrapper extraction artifact (CAGR 11.5% ⇒ real trades).

**39. composite-c2-equityfactor: regime-robustness split — the 0.574 is bull-market-concentrated (pro-cyclical long-only), not regime-agnostic; capital-preserving but unprofitable in the 2022 bear (2026-06-23)**

a 4-regime sub-period split of the strongest COMP backbone (composite-c2-equityfactor, full-window Sharpe 0.574 / 1761 tradeable dates, #34) on an IBKR-fee clone (`1630-robustness-c2-regime`, project 33280244; clone main.py = canonical + date parameterization only, all other logic byte-identical) decomposes the aligned 2018-2024 window into four market regimes — R1 2018-19 pre-COVID normal, R2 2020-21 COVID bull, R3 2022 bear/rate-hike, R4 2023-24 recovery/AI. The split partitions the window **exactly** (R1 503 + R2 505 + R3 251 + R4 502 = 1761 = full window, no overlap/gap).

**Result — sharply bimodal, NOT uniform**: R2 COVID-bull **Sharpe 1.040 / CAGR 18.9% / PSR 50.5%** and R4 recovery **Sharpe 1.094 / CAGR 22.6% / PSR 82.7%** are where c2 earns *all* its edge (both statistically meaningful; both US-large-cap bull/recovery markets — the 2020-21 stimulus rally and the 2023-24 AI/Mag7 rally); R1 pre-COVID normal is weak (**Sharpe 0.131 / PSR 17.1%**) and R3 2022 bear is **negative (Sharpe −0.219 / CAGR −0.7%)** — c2 *loses money* in the rate-hike bear. The full-window 0.574 is thus a blend dominated by the two bull/recovery regimes (~57% of dates at Sharpe ~1.07) offsetting the weak/negative normal+bear regimes (~43% at ~0.0).

**This nuances finding #34's "genuinely robust rather than period-overfit"**: c2 is indeed robust to *fee/window alignment* (the #1630 discriminator — it holds at 0.574 with real IBKR fees, doesn't collapse on alignment like FamaFrenchAllWeather #35), but it is **NOT regime-agnostic** — it is a pro-cyclical long-only US-large-cap factor strategy whose Sharpe is a bull-market premium, exactly as expected for its design (Value/Quality/LowVol/Momentum on FineFundamental top-25 market-cap).

The defensible nuance: c2's risk model (SectorCapRiskModel 18% portfolio-DD circuit breaker + 4% trailing stops) **does protect capital in the bear** — the 2022 MaxDD is only 9.5% (well under the 18% cap), so c2 *survives* downturns without profiting from them, a meaningful property for a long-only backbone.

**Contrast with EMATrend (#36)**: EMATrend's caveat is *survivorship* (Mag7-universe bias); c2's caveat here is *regime-concentration* (pro-cyclical) — but c2's *constitution* is cleaner (factor-diversified across 25 stocks, no Mag7 concentration), so the "most robust COMP constitution" claim (#34) stands *constitutionally* even as the Sharpe is regime-concentrated.

**Practical takeaway**: do not over-read the 0.574 as a transferable all-weather edge — it is a large-cap-US bull/recovery premium with capital preservation in bears, not a regime-agnostic alpha.

Backtests R1 `63605ad6` / R2 `7fb12884` / R3 `1342b6b5` / R4 `7dc742af` (clone project 33280244, compile `614c4d79`; canonical c2 project 32981222 / full-window bt `8eecba32` unchanged). Same totalOrders=0 wrapper extraction artifact (regime CAGRs 4.8%/18.9%/−0.7%/22.6% ⇒ real trades). `See #1630`.

**40. PCA-StatArbitrage: catalog 0.40 does not survive alignment (0.399 → 0.205, PSR 1.4% noise) — and the drop is a WINDOW-effect, not fee (fee-IMMUNITY proven by a hardcoded no-brokerage clone) (2026-06-23)**

PCA statistical-arbitrage mean-reversion (sklearn PCA + per-stock OLS on log prices → residual z-scores → contrarian long the z<−1.5 tail, weights ∝ z-deviation, full monthly rotation via `set_holdings(liquidate=True)`; top-100 US equities by dollar volume, coarse-universe monthly refresh, IBKR margin; source: Hands-On AI Trading Ch.06 Ex.13 / Avellaneda-Lee 2010) on 2018-2025 gives Sharpe **0.205**, CAGR 6.762%, MaxDD 31.8%, PSR **1.418%** (noise-level), 2011 tradeable dates.

The **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: catalog 0.40 (README 0.399) does not survive alignment — a −49% drop, PSR 1.4% = indistinguishable from noise.

Two findings, both correcting a pre-run prediction:

**(a) the drop is a WINDOW-effect, NOT a fee-effect** — the catalog README 0.399 was already an IBKR number (the canonical `main.py` carries `set_brokerage_model(IBKR)`), so the 0.399→0.205 degradation cannot be fee-remediation; it is the aligned window's 2024-25 Mag7-momentum melt-up, structurally hostile to contrarian mean-reversion (the "oversold" residual names keep falling, the winners keep winning, so the z<−1.5 contrarian tail bleeds).

**(b) Fee-IMMUNE — refuting the "monthly full-rotation ⇒ fee-vulnerable" prediction** — a hardcoded no-brokerage clone of the identical logic (project 33281920, backtest `63fce57d`) returns byte-identical Sharpe 0.205 / CAGR 6.762% / MaxDD 31.8% / PSR 1.418%; physically removing the brokerage changes nothing at 3-decimal precision. The pre-run hypothesis was that the monthly full-rotation of a *changing* contrarian subset would make this the archetypal fee-collapse case; the result instead **extends the #1630 discriminator regime-3 near-immune class to PCA stat-arb** — the fee-HOMOGENEITY of the US-equity top-100 basket (<0.25 bps/trade via IBKR, cf EMA-Cross-Stocks IBKR 0.991 = no-brokerage 0.991, fee-finding #2 at line 740) dominates over the per-event turnover size, exactly as for EMA-CS / ML-RandomForest / composite-c2 (#34). The discriminator is basket fee-homogeneity × realized turnover, not mechanism (mean-reversion vs trend) or rotation frequency in isolation.

Net honest headline: "PCA-StatArb = 0.205 on alignment, PSR 1.4% noise, and it is the aligned WINDOW that broke it, not the fees."

Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified, un-robust). Backtest `9169c962` (IBKR) / `63fce57d` (no-fee control), project 33281920 (baseline-clone `1630-baseline-PCAStatArbitrage`, created fresh; only main.py date-aligned to 2018-2025, IBKR unchanged = catalog standard). Same totalOrders=0 wrapper extraction artifact (CAGR 6.76% ⇒ real trades). Numbered #40 assuming #4089 (c2-regime) takes #39 on rebase post-#4084-merge.

## #1630 Aligned Backbone Leaderboard (2018-2025, no-ML)

The 15 IND/COMP/RISK/FACTOR Tier-4 candidates verified across Key-findings #22–#36, ranked by aligned Sharpe. The three Tier-3 negatives (#24 / #27 / #28) document strategies whose published headline Sharpe does not survive the modern, fee-aware window; none of the 12 Tier-2 positives is statistically significant (PSR < 30 %) — no no-ML backbone reaches the Tier-1 > 0.5 robustness bar on the aligned period, though EMATrend (0.611) and composite-c2-equityfactor (0.574) come closest.

| Rank | Strategy | Family | Sharpe | CAGR % | MaxDD % | PSR % | Tier | Finding |
|------|----------|--------|-------:|-------:|--------:|------:|------|:--------|
| 1 | Framework_Composite_EMATrend | COMP | **0.611** | 16.67 | 27.9 | 19.8 | 2 | #36 (Mag7 survivorship caveat) |
| 2 | composite-c2-equityfactor | COMP | **0.574** | 11.94 | 18.6 | 25.8 | 2 | #34, #39 (most robust constitution; regime-bimodal — bull/recovery-concentrated, capital-preserving in bear) |
| 3 | HAR-RV-J-Kelly | RISK | 0.524 | 14.08 | 37.1 | 10.7 | 2 | #29 |
| 4 | GlobalMacro-Regime | RISK | 0.454 | 9.80 | 22.8 | 16.7 | 2 | #26 |
| 5 | FamaFrench | FACTOR | 0.445 | 11.11 | 24.1 | 11.9 | 2 | #32 |
| 6 | Framework_Composite_FamaFrenchAllWeather | COMP | 0.338 | 6.58 | 13.1 | 22.9 | 2 | #35 (period-overfit collapse) |
| 7 | Vol-GARCH-Target | RISK | 0.325 | 6.97 | 10.8 | 14.9 | 2 | #30 (tightest MaxDD) |
| 8 | Vol-Ensemble-Conservative | RISK | 0.265 | 6.14 | 10.4 | 13.6 | 2 | #31 (tightest MaxDD) |
| 9 | composite-c1-multiasset | COMP | 0.258 | 6.49 | 17.0 | 8.7 | 2 | #33 |
| 10 | AssetClassMomentum-QC | IND | 0.220 | 6.64 | 28.1 | 3.8 | 2 | #22 |
| 11 | Cloud-VolTargeting (v1) | RISK | 0.207 | 6.72 | 38.2 | 2.4 | 2 | #25 |
| 12 | Cloud-RiskParity-Composite | RISK | 0.027 | 3.50 | 24.4 | 1.2 | 2 | #23 |
| 13 | Cloud-SectorRotation-Momentum | IND | -0.029 | 2.13 | 42.7 | 0.5 | 3 | #24 |
| 14 | TermStructureCommodities-QC | IND | -0.244 | -31.48 | 96.8 | 0.0 | 3 | #28 (catastrophic) |
| 15 | MomentumRegime-AdaptiveWeights | COMP | -0.729 | 1.88 | 4.3 | 17.4 | 3 | #27 (double-defense) |

**Read-across**: (i) the COMP family spans the entire range (0.611 → -0.729) — the component alpha / PCM choice dominates the framework wiring (#33 / #34 share one scaffold yet differ 2×); (ii) the RISK family's edge is drawdown control rather than return — Vol-GARCH-Target and Vol-Ensemble-Conservative deliver the tightest MaxDD of the campaign (10.4-10.8 %) at a modest Sharpe (#30 / #31); (iii) every positive is statistically non-significant (PSR < 30 %) — no no-ML backbone clears Tier-1 on the aligned window; (iv) the three negatives (#24 / #27 / #28) confirm that published headline Sharpe figures routinely fail to survive alignment to a modern fee-aware period.

## Residual multi-asset pool grounding (#1630 deep-queue, 2026-06-24)

The two remaining unvetted multi-asset rotation strategies from the #1630 pool — **RiskParity (#46, catalog 0.40)** and **DualMomentum (#48, catalog 0.35)** — were verified on the aligned 2018-2024 window under native IBKR margin brokerage (real fees), closing the ai-01 deep-queue item "RiskParity/DualMomentum" (BTC-ML remains deferred to multi-seed #6). Both were deployed as fresh aligned baseline-clones (33286158, 33286487) with only the start/end dates changed to the methodology window; the strategy logic, universe and IBKR brokerage are byte-identical to the canonical `projects/*/main.py`. Daily US-ETF data is free on QC Cloud → 0 QCC consumed.

**RiskParity (#46) — true inverse-volatility risk-parity (Sharpe 0.361, PSR 11.3%).** This is the genuine inverse-volatility risk-parity implementation (rank-based RP "v4 BEST", docstring 0.515): SMA200 trend filter + 12m-momentum rank × close/std60 inverse-volatility weighting, BND safe-haven when fewer than 2 assets trend, monthly rebalance over SPY/EFA/GLD/DBC/TLT. The aligned 0.361 is **−30% versus the student 0.514** (project 31872286, *hardcoded 2015 start, non-aligned) — the regime-4 fee-friction signature: a multi-asset basket carrying bonds (TLT/BND) and commodities (GLD/DBC) sleeves is **not fee-homogeneous**, exactly the AllWeather −30% collapse pattern. It is nonetheless the **best risk-parity-family member verified**: far above the mis-named equal-weight Cloud-RiskParity-Composite (0.027, #23) and AssetClassMomentum (0.22, #22), confirming that true inverse-volatility weighting dominates both naive equal-weight rotation and simple top-N momentum within this family. PSR 11.3% remains non-significant — no Tier-1 robustness.

**DualMomentum (#48) — Antonacci canonical dual-momentum (Sharpe 0.350, PSR 9.6%).** The canonical Antonacci (2014) construction: a dual absolute filter (price > SMA200 **and** 6m return > 0) gates a relative 12m-return ranking over **all** passing candidates with momentum-tilt weighting, BND safe-haven below the minimum-trending threshold, monthly rebalance over SPY/EFA/EEM/TLT/GLD/DBC. The aligned 0.350 is **essentially flat versus the catalog 0.35** — no fee-collapse, the catalog was already honest. Notably it delivers the **tighter MaxDD (14.9%)** of the two siblings (versus RiskParity 18.4%), the dual absolute+relative filter plus the BND refuge buying genuine drawdown control. The headline caveat is statistical: the student DualMomentum backtest (project 31798582) reported Sharpe 0.493 / **PSR 54.9%** but on only a 2-year hardcoded window (2023-2025) — that 54.9% was a **small-sample artifact** that does not survive the full aligned window (PSR 9.6%), precisely the FamaFrenchAllWeather 87.5%→22.9% collapse pattern (finding #35). PSR signficance shrinks as `√T` grows; a 2-year PSR > 50% is not transferable evidence of an edge.

**Family conclusion — the multi-asset momentum/risk-parity family is grounded.** The two siblings land within 0.01 of each other on the aligned window (0.361 / 0.350), both regime-4 fee-vulnerable, both non-significant (PSR 9.6-11.3%). Together with the already-vetted Cloud-RiskParity-Composite (0.027), AssetClassMomentum-QC (0.22), GlobalMacro-Regime (0.454, #26 — the best of this family, which adds an SPY regime switch), and the composite-c1-multiasset (0.258, #33), the family forms a coherent regime-4 band: **survives real fees at a modest Sharpe (0.25-0.46) but never reaches Tier-1 significance on the aligned window.** The #1630 realized-turnover discriminator holds — multi-asset rotation through bonds/commodities/EM sleeves incurs fee-friction that the fee-homogeneous US-large-cap equity strategies (regime-3 near-immune) avoid. The honest verdict is "verdict faible-valeur = grounding légitime": these are capital-preserving, drawdown-controlled rotation strategies, not robust alpha generators, and the higher student PSRs were window artifacts. This closes the residual multi-asset pool of the #1630 deep-queue.

**Residual-pool + multi-asset-rotation family, ranked by aligned Sharpe (all regime-4 fee-vulnerable):**

| Strategy | Mechanism | Aligned Sharpe | MaxDD % | PSR % |
|----------|-----------|---------------:|--------:|------:|
| GlobalMacro-Regime (#26) | inv-vol RP + SPY regime switch | 0.454 | 22.8 | 16.7 |
| RiskParity (#46) | inv-vol RP + BND safe-haven | 0.361 | 18.4 | 11.3 |
| DualMomentum (#48) | Antonacci dual-mom + BND | 0.350 | 14.9 | 9.6 |
| AssetClassMomentum-QC (#22) | top-3 ETF momentum | 0.220 | 28.1 | 3.8 |
| Cloud-RiskParity-Composite (#23) | equal-weight momentum (mis-named) | 0.027 | 24.4 | 1.2 |

The table makes the family hierarchy explicit: the regime switch (#26) is the differentiator that lifts GlobalMacro-Regime above the pure-rotation siblings, and the two newly-vetted strategies (#46/#48) sit in the middle of the band — better than naive rotation, worse than the regime-aware variant, all non-significant.

## Comparison: Best-vs-Aligned

| Strategy | Best Sharpe | Aligned Sharpe | Delta | Diagnostic |
|----------|------------|----------------|-------|------------|
| EMA-Cross-Alpha | 0.996 | -0.010 | -1.006 | Period overfitting severe |
| EMA-Cross-Stocks | 0.87 | 0.891 | +0.021 | Performance confirmee, ameliore |
| TrendStocks-Alpha | 0.609 | 0.519 | -0.090 | Legere degradation |
| AllWeather | 0.67 | 0.631 | -0.039 | Stable, confirme |
| MomentumStrategy | 0.57 | 0.555 | -0.015 | Performance confirmee |
| TrendFollowing | ~0.8 | 1.072 | +0.272 | Ameliore sur longue periode |
| Crypto-MultiCanal | ~0.6 | 0.581 | ~0 | Performance confirmee |
| VolTarget-Momentum | ~0.65 | 0.648 | ~0 | Performance confirmee |
| PuppiesOfTheDow-QC | 1.99 | 0.302 | -1.688 | Period overfitting severe (catalog = fenetre glissante 12 ans) |
| LeveragedETFMomentum-QC | 1.80 | 1.779 | -0.021 | Performance confirmee (mais MaxDD 53%) |
| Framework_Composite_TrendWeather | 1.16 | 0.948 | -0.212 | Legere degradation, composite robuste |
| HighBookToMarketFScore-QC | 2.09 | 0.411 | -1.679 | Period overfitting severe + MaxDD 60% |
| MeanReversion | 0.29 (old) | **0.810** (v5.2) | +0.520 | v5.2 IBKR dramatically better. Calmar 1.34 |
| AdaptiveAssetAllocation | untested | 0.509 | +0.509 | First aligned baseline. Min-var + momentum |
| PairsTrading | -0.36 | -0.280 | +0.080 | Marginal improvement, still exploratoire |

---

## Next Steps

1. ~~**Standardized backtest period**: Re-run all 62 tested + 39 untested strategies on 2018-01-01 → 2024-12-31~~ — Done, 25 baselines verified via QC Cloud API (See #1630)

2. ~~**Run aligned baselines for AllWeather/SectorMomentum/EMA-Cross-Stocks/MomentumStrategy**~~ — Done, all 4 re-backtested via QC Cloud

3. ~~**Student strategies (ESGF #1405)**: DualMomentum, RiskParity, ValueFactor, OptionWheel backtestees~~ — Done, 4/6 valides

3b. ~~**Run baselines for MeanReversion, AAA, PairsTrading**~~ — Done 2026-06-11. MeanReversion promoted Tier 2→1 (0.81), AAA promoted Tier 4→1 (0.509), PairsTrading confirmed exploratoire (-0.28)

4. ~~**Transaction cost sensitivity analysis**: Estimated turnover and cost impact for all 10 research baselines~~ — Done (See #1407)

5. ~~**Transaction cost re-backtest**: Add `SetBrokerageModel` + configurable brokerage parameter~~ — Done, #2575 + fee sweep EMA-Cross-Stocks + Crypto-MultiCanal (See #2471, #2575, #2588)

6. **Cross-seed validation (gates #7)**: ≥4 seeds (0/1/7/42/99) for ML/DL/RL strategies — multi-cycle, requires re-training each model on ≥4 seeds and re-backtesting (heavy QC + GPU). **Blocks #7**: σ_cross_seed is a prerequisite input for Edge vs σ. See section "Edge vs σ — statut & dépendance" below.

7. **Edge vs σ (gated on #6)**: Compute `(Sharpe - baseline_Sharpe) / σ_cross_seed` for ML/DL/RL multi-seed strategies vs B&H baseline. **Not computable standalone** — requires σ_cross_seed from #6. For non-ML (IND/COMP/RISK/OPT) strategies there is no σ_cross_seed (single-run), so Edge vs σ applies only to the ML/DL/RL subset once #6 delivers. See section "Edge vs σ — statut & dépendance" below.

8. **Trend-Following repo/cloud drift**: repo code gives Sharpe 0.365 on 2018-2024 vs published 1.072 (2018-2025, prior cloud state) — identify which code version produced 1.072 and align repo (see Key finding 18)

9. **No-ML backbone campaign complete (2026-06-23)**: all IND/COMP/RISK/FACTOR Tier-4 candidates verified (Key-findings #22–#36, 15 baselines). Remaining Tier-4 = OPT (#73/#75, naked-options MaxDD > 100 % simulator caveat) + ML/DL/RL (#63–#71/#76, deferred to #6 multi-seed / training-specialist). Research-Executor (#92) is a research harness, not a strategy — out of baseline scope (Key-finding #37).

---

## Edge vs σ — statut & dépendance (#1630 items #6 → #7)

**Définition** (spec #1630) : `Edge_vs_σ = (Sharpe_strategy − baseline_Sharpe) / σ_cross_seed`, où `baseline_Sharpe` = Sharpe d'un benchmark buy-and-hold (SPY pour US equities, BTC pour crypto) sur la même période, et `σ_cross_seed` = écart-type du Sharpe de la stratégie across ≥4 seeds d'entraînement.

**Statut : non calculable ce cycle — blocker structurel documenté.**

### Dépendance : #7 exige #6 (σ_cross_seed)

L'Edge vs σ **ne peut pas être calculé sans `σ_cross_seed`**, qui n'existe que pour les stratégies **multi-seed** — c'est précisément ce que l'item #6 (cross-seed validation) doit produire. Donc :

- **#7 est gated on #6** : tant que #6 n'a pas livré ≥4 seeds par stratégie ML/DL/RL, il n'y a pas de `σ_cross_seed` en entrée.
- **#6 est multi-cycle/lourd** : re-entraîner chaque modèle ML/DL/RL sur ≥4 seeds (0/1/7/42/99) + re-backtester chaque run consomme massivement le rate-limit QC (10 backtests/min cluster) et/ou le GPU. Pas un grain atomique 1-cycle — découper en sous-lots par famille (ML equity, DL, RL) quand priorisé.

### Périmètre : ML/DL/RL uniquement

Pour les stratégies **non-ML** (IND/COMP/RISK/OPT/STAT), il n'y a **pas de σ_cross_seed** : elles sont déterministes (single-run, pas de seed d'entraînement). L'Edge vs σ spec #1630 vise donc **explicitement les ML/DL/RL** (« Edge vs σ calculé pour les strategies ML/DL/RL multi-seed »). Les stratégies non-ML reçoivent un edge simplifié `(Sharpe − baseline_Sharpe)` en σ-unités via la volatilité réalisée si besoin, mais pas un edge cross-seed.

### Stratégies éligibles (une fois #6 livré)

Les entrées ML/DL/RL du catalogue, par tier :
- **Tier 1 (robuste)** : LongShortHarvest, Positive-Negative-Splits-ML, DynamicVIXSpyRegime, MacroFactorRotation, Portfolio-Optimization-ML, CausalEventAlpha, Gaussian-Direction-Classifier, ML-Temporal-CNN, ML-RandomForest, ML-Trend-Scanning, ML-FeatureEngineering, Markov-Regime-Detection, ML-XGBoost, RegimeSwitching, Temporal-CNN-Prediction, RL-DQN-Trading, LSTM-Forecasting.
- **Tier 2/3/4** : ML-Classification, ML-Regression, ML-Ensemble, ML-DeepLearning, DL-LSTM, RL-Portfolio, Reinforcement-Learning-Trading, BTC-ML, Crypto-LSTM-Prediction, ML-Reversion-Trending, ML-Chronos-Foundation, Chronos-Foundation-Forecasting, Adaptive-Conformal-Risk, ML-Gaussian-Classifier, ML-TextClassification, ML-EnhancedPairs, PCA-StatArbitrage.

### Baseline_Sharpe (B&H, déjà mesurable)

La moitié `baseline_Sharpe` de la formule **est** calculable dès maintenant (B&H sur la période aligned) — mais publiera un edge complet seulement après #6 :

| Benchmark | Période | B&H Sharpe (approx.) | Note |
|-----------|---------|----------------------|------|
| SPY (US equities) | 2018-2025 | ~1.15 | Barre élevée — une stratégie equity qui ne bat pas SPY B&H n'a pas d'edge |
| BTC (crypto) | 2020-2025 | ~1.0 (bull) | Crypto bull-market caveat — cf Portfolio-IBKR-Coinbase-Hybrid |

**Note honnête** : un edge simplifié `(Sharpe_strategy − Sharpe_B&H)` (sans normalisation σ_cross_seed) est trompeur — il ne distingue pas une stratégie avec un edge stable d'une avec un edge bruité. La normalisation par σ_cross_seed est précisément ce qui fait la valeur de la métrique. D'où le gate #6.

### Conclusion

#7 n'est pas un grain livrable ce cycle. Le chemin correct : **#6 d'abord** (sous-lots par famille ML/DL/RL, multi-cycle), puis #7 se calcule trivialement (une division par ligne). Cette section documente la dépendance pour qu'elle soit traçable plutôt qu'un vague « à calculer ». Si une priorisation est voulue, starter pack suggéré : **les 4 true leaders PSR>50%** (Positive-Negative-Splits-ML 82.3%, Framework_Composite_TrendWeather 77.9%, FamaFrenchAllWeather 87.5%, Portfolio-IBKR-Coinbase-Hybrid 62%) — confirmer leur edge en σ-unités cross-seed est le plus utile pédagogiquement.

---

## Transaction Cost Sensitivity (See #1407)

Source code analysis of all 10 research projects. Zero strategies have explicitly configurable transaction cost parameters (`SetFeeModel` / custom fee). Costs are implicit in the brokerage model or use QC defaults.

### Fee Model Configuration

| Project | QC ID | `SetFeeModel` | `SetBrokerageModel` | Configurable | Default Fees |
|---------|-------|---------------|----------------------|--------------|--------------|
| Trend-Following | 28797562 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| EMA-Cross-Stocks | 28789946 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| MomentumStrategy | 28657837 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| AllWeather | 28657833 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| VolTarget-Momentum | 30784745 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| Crypto-MultiCanal | 30750734 | NONE | `BINANCE, CASH` | YES (`brokerage=binance/none`) | Binance schedule (0.1%) |
| Portfolio-IBKR-Binance | 31717642 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| MomentumRegime | 31243821 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| TrendStocks-Alpha | 28885507 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| EMA-Cross-Alpha | 28885488 | NONE | `IBKR, MARGIN` | YES (`brokerage=ibkr/none`) | IBKR tiered |
| ForexCarry | 28657908 | NONE | `OANDA, MARGIN` | YES (`brokerage=oanda/none`) | OANDA schedule |
| AdaptiveAssetAllocation | 28693649 | NONE | NONE | NO | QC defaults |
| MeanReversion | 30776121 | NONE | `IBKR, MARGIN` | NO | IBKR tiered (v5.2 IBKR baseline) |
| PairsTrading | 28693651 | NONE | NONE | NO | QC defaults |

### Turnover Estimation

| Project | Rebalance Freq | Universe | Est. Annual Trades | Turnover | Cost Sensitivity |
|---------|----------------|----------|--------------------|----------|------------------|
| TrendFollowing | Monthly | 7 ETFs | ~24-48 | MEDIUM | MEDIUM |
| EMA-Cross-Stocks | Daily (5% drift) | 5 stocks | ~125-250 | HIGH | **HIGH** |
| MomentumStrategy | Monthly + stop-loss | 11 sectors | ~48-96 | MEDIUM-HIGH | HIGH |
| AllWeather | Quarterly (3% drift) | 4 ETFs | ~8-16 | LOW | LOW |
| VolTarget-Momentum | Monthly (leverage 0.3-1.5x) | 7 ETFs | ~24-48 | MEDIUM | MEDIUM |
| Crypto-MultiCanal | Daily (signal) | 1 (BTC) | ~50-100 | HIGH | **HIGH** (0.1% crypto) |
| Portfolio-IBKR-Binance | One-shot | 3 assets | ~3 | VERY LOW | NEGLIGIBLE |
| MomentumRegime | Monthly | 4 assets | ~24-48 | MEDIUM | MEDIUM |
| TrendStocks-Alpha | Weekly | 15 stocks | ~150-300 | HIGH | **HIGH** |
| EMA-Cross-Alpha | Daily (1-day insight) | 5 stocks | ~125-250 | HIGH | **HIGH** |

### Cost Sensitivity Ranking

**TIER 1 — Highly Sensitive** (cost change would significantly impact Sharpe):

1. **EMA-Cross-Stocks** — Daily rebalance, 5 stocks, 5% drift threshold. No brokerage model = default QC fees. A 2x fee increase could eat 50-100bps CAGR.
2. **EMA-Cross-Alpha** — Daily insight emission, IBKR margin. Same high-frequency cross logic.
3. **TrendStocks-Alpha** — 15-stock universe, weekly rebalance. Largest universe = most trades.
4. **Crypto-MultiCanal** — Daily BTC signals, Binance CASH (already has real 0.1% crypto fees). Crypto percentage-based fees amplify impact vs per-share equity fees.
5. **MomentumStrategy** — Monthly rotation of top-4 from 11 sectors + daily stop-loss checks. Stop-losses generate unplanned trades.

**TIER 2 — Moderately Sensitive:**

6. **TrendFollowing** — Monthly rebalance, 6 risky + 1 safe. Moderate turnover on regime changes.
7. **VolTarget-Momentum** — Monthly with leverage scaling (0.3x to 1.5x). Leverage changes mean larger position sizes.
8. **MomentumRegime** — Monthly composite on 4 assets, IBKR margin. Moderate.

**TIER 3 — Low Sensitivity:**

9. **AllWeather** — Quarterly + 3% drift threshold, 4 static-weight ETFs. Very few trades.
10. **Portfolio-IBKR-Binance** — One-shot buy of SPY/BND/AAPL. 3 trades total. Zero sensitivity.

### Recommended Actions

| Priority | Project | Action | Rationale |
|----------|---------|--------|-----------|
| HIGH | EMA-Cross-Stocks | Add configurable fee + 0x/1x/2x sweep | Highest turnover, no brokerage = blind to real costs |
| HIGH | TrendStocks-Alpha | Same | Largest universe, weekly rebalance |
| HIGH | Crypto-MultiCanal | Test 0x/2x Binance fees | **Done** — Binance Sharpe 0.333 vs None 0.181 (See #2590) |
| MEDIUM | MomentumStrategy | Add fee sweep | Stop-losses create unplanned trades |
| MEDIUM | VolTarget-Momentum | Add fee sweep | Leverage amplifies trade sizes |
| LOW | AllWeather | Optional | Very few trades, minimal cost impact |
| SKIP | Portfolio-IBKR-Binance | Not applicable | One-shot, 3 trades total |

### Fix Pattern (for fee sweep)

```python
# In initialize(), add configurable fee model
self.transaction_cost_bps = self.get_parameter("transaction_cost_bps", 5)  # default 5bps
for ticker in self.all_tickers:
    security = self.add_equity(ticker, Resolution.DAILY)
    security.set_fee_model(ConstantFeeModel(self.transaction_cost_bps / 10000 * self.portfolio.total_portfolio_value * 0.20))
```

### Estimated Fee Impact Analysis

**Methodology**: Each strategy's annual turnover is translated into a round-trip cost drag. Fee assumptions: US equities/ETFs = 5 bps one-way (10 bps round-trip, IBKR pricing), Crypto = 10 bps one-way (20 bps round-trip, Binance), FX = 2 bps one-way. QC-reported CAGR is treated as near-zero-fee (no explicit `SetFeeModel` in 8/10 strategies).

#### Annual Cost Drag

| Project | Est. Turnover | Round-trip Fee | Annual Cost Drag | CAGR% (QC) | Est. Net CAGR% | CAGR Erosion |
|---------|---------------|----------------|------------------|------------|----------------|--------------|
| TrendFollowing | 2x | 10 bps | 20 bps (0.20%) | 23.2 | 23.0 | -20 bps |
| EMA-Cross-Stocks | 5x | 10 bps | 50 bps (0.50%) | 26.2 | 25.7 | -50 bps |
| MomentumStrategy | 3x | 10 bps | 30 bps (0.30%) | 11.7 | 11.4 | -30 bps |
| AllWeather | 0.5x | 10 bps | 5 bps (0.05%) | 9.0 | 9.0 | -5 bps |
| VolTarget-Momentum | 2.5x | 10 bps | 25 bps (0.25%) | 14.7 | 14.5 | -25 bps |
| Crypto-MultiCanal | 4x | 20 bps | 80 bps (0.80%) | 8.2 | 7.4 | -80 bps |
| Portfolio-IBKR-Binance | 0.15x | 10 bps | 1.5 bps (0.02%) | 15.7 | 15.7 | -2 bps |
| MomentumRegime | 2x | 10 bps | 20 bps (0.20%) | 4.7 | 4.5 | -20 bps |
| TrendStocks-Alpha | 6x | 10 bps | 60 bps (0.60%) | 15.9 | 15.3 | -60 bps |
| EMA-Cross-Alpha | 5x | 10 bps | 50 bps (0.50%) | 2.8 | 2.3 | -50 bps |

**Calculation**: Annual Cost = Turnover x Round-trip Fee. Example: EMA-Cross-Stocks = 5x x 10 bps = 50 bps/year drag.

#### Sharpe Ratio Erosion

| Project | Aligned Sharpe | Est. Vol% | Annual Cost | Sharpe Erosion | Est. Net Sharpe | Erosion % |
|---------|---------------|-----------|-------------|----------------|-----------------|-----------|
| TrendFollowing | 1.072 | 21.6% | 20 bps | -0.009 | 1.063 | -0.9% |
| EMA-Cross-Stocks | 0.891 | 29.4% | 50 bps | -0.017 | 0.874 | -1.9% |
| MomentumStrategy | 0.555 | 21.1% | 30 bps | -0.014 | 0.541 | -2.5% |
| AllWeather | 0.631 | 14.3% | 5 bps | -0.003 | 0.628 | -0.5% |
| VolTarget-Momentum | 0.648 | 22.7% | 25 bps | -0.011 | 0.637 | -1.7% |
| Crypto-MultiCanal | 0.581 | 14.1% | 80 bps | -0.057 | 0.524 | **-9.8%** |
| Portfolio-IBKR-Binance | 0.519 | 30.3% | 1.5 bps | -0.001 | 0.519 | -0.1% |
| MomentumRegime | 0.185 | 25.4% | 20 bps | -0.008 | 0.177 | -4.3% |
| TrendStocks-Alpha | 0.519 | 30.6% | 60 bps | -0.020 | 0.499 | -3.8% |
| EMA-Cross-Alpha | -0.010 | 28.0% | 50 bps | -0.018 | -0.028 | N/A |

**Calculation**: Vol = CAGR / Sharpe. Sharpe erosion = Cost / Vol. Example: Crypto-MultiCanal = 0.80% / 14.1% = 0.057 Sharpe erosion.

#### Fee Resilience Ranking

| Rank | Project | Cost Drag | Sharpe Erosion | Break-even Fee (one-way) | Resilience |
|------|---------|-----------|----------------|--------------------------|------------|
| 1 | Portfolio-IBKR-Binance | 1.5 bps | 0.5 bps | ~52,000 bps | Near-immune |
| 2 | AllWeather | 5 bps | 3 bps | 1,800 bps | Very high |
| 3 | TrendFollowing | 20 bps | 9 bps | 580 bps | High |
| 4 | VolTarget-Momentum | 25 bps | 11 bps | 294 bps | High |
| 5 | MomentumStrategy | 30 bps | 14 bps | 195 bps | High |
| 6 | EMA-Cross-Stocks | 50 bps | 17 bps | 262 bps | Moderate |
| 7 | TrendStocks-Alpha | 60 bps | 20 bps | 133 bps | Moderate |
| 8 | MomentumRegime | 20 bps | 8 bps | 118 bps | Moderate |
| 9 | Crypto-MultiCanal | 80 bps | 57 bps | 103 bps | **Vulnerable** |
| 10 | EMA-Cross-Alpha | 50 bps | 18 bps | 28 bps | **Vulnerable** |

**Break-even fee** = CAGR / Turnover. Below this threshold the strategy remains net profitable.

#### Key Takeaways

1. **Crypto fees are NOT the primary risk factor (revised)**: Crypto-MultiCanal fee sweep (backtest IDs `9ad550e9`, `56d54a3c`) shows Binance (real 0.1% fees) Sharpe 0.333 vs no-brokerage Sharpe 0.181. Fees actually improve Sharpe by +0.152 (+84%), likely because the brokerage model's cash constraints filter low-quality trades. The theoretical -9.8% erosion estimate was wrong — real backtesting contradicts it.
2. **EMA-Cross-Stocks: fees negligible** (backtest IDs from #2588): IBKR Sharpe 0.991 vs no-brokerage Sharpe 0.991 (identical). US equity fees via IBKR are <0.25 bps per trade — negligible impact even at high turnover.

### Verified Fee Sweep Results

Backtest-validated fee sensitivity for strategies with configurable brokerage parameter.

| Project | With Fees | Sharpe (fees) | Sharpe (no fees) | Delta | Verdict |
|---------|-----------|---------------|-------------------|-------|---------|
| EMA-Cross-Stocks | IBKR Margin | 0.991 | 0.991 | 0.000 | **Near-immune** (US equity fees negligible) |
| Crypto-MultiCanal | Binance Cash | 0.333 | 0.181 | **+0.152** | **Fees improve** (cash constraints filter bad trades) |

**Key insight**: Theoretical fee erosion estimates can be wrong. Binance CASH account type enforces cash settlement rules that prevent over-leveraging, improving risk-adjusted returns. Real backtesting is essential.
2. **EMA-Cross-Alpha is fragile on multiple dimensions**: Already negative Sharpe, thin fee margin (28 bps break-even), and high turnover. Confirms "exploratoire" classification.
3. **High-turnover equity strategies lose meaningful CAGR**: EMA-Cross-Stocks (-50 bps) and TrendStocks-Alpha (-60 bps) each lose ~0.5-0.6% annually. Warrant explicit `SetFeeModel`.
4. **No profitable strategy flips unprofitable** at realistic fees. The risk is Sharpe degradation, not sign flip.
5. **Slippage excluded**: Market impact on small-caps (EMA-Cross-Stocks) could add 5-15 bps/trade, doubling effective cost for those strategies.

---

## Data Source

- QC Cloud API via `qc-mcp-lite` — backtest IDs verified 2026-06-05, updated 2026-06-11 (MeanReversion, AAA, PairsTrading)
- `projects/catalog.json` — 114 entries, authoritative metadata
- `projects/README.md` + `STRATEGIES_DETAIL.md` — human-readable catalog