High-Sharpe Multi-Alpha Engineering & Convex Optimization Engine
A rigorous classical statistics approach to alpha generation, risk modeling, and advanced portfolio construction yielding a 2.51 Sharpe Ratio across 21 years of out-of-sample stress-tested data.
AlphaStack is a professional-grade quantitative research framework designed to discover, combine, and optimize multi-alpha signals into market-resilient portfolios. Bypassing black-box machine learning in favor of pure classical statistics, this engine implements a rigorous pipeline of signal generation (cross-sectional momentum, short-term reversal, residual momentum, IVOL, BAB, etc.), Information Coefficient (IC) weighting, and seven distinct portfolio construction engines.
The crowning strategy combines Maximum Diversification, Volatility Targeting, and a dynamic VIX Circuit Breaker to deliver state-of-the-art risk-adjusted returns while effectively capping tail-risk drawdowns.
The framework operates on a modular, walk-forward pipeline executing weekly rebalances over an extensive 21-year (2004–2024) horizon with strict Out-of-Sample (OOS) validation spanning into 2026.
graph TD;
A[Data Pipeline <br/> yFinance, Fama-French, VIX] --> B(Alpha Generation <br/> 13 Classical Signals);
B --> C{Signal Combination <br/> IC-Weighting & Shrinkage};
C --> D[Portfolio Optimization <br/> CVXPY, HRP, MaxDiv];
D --> E[Risk Management <br/> Vol-Targeting & Circuit Breaker];
E --> F((Performance & Statistical <br/> Inference Verification));
Generates industry-neutralized, cross-sectionally ranked signals evaluated via Spearman rank correlation (IC):
- Momentum: 12-1, 6-1, 3-1, 12-7 Intermediate (Jegadeesh-Titman 1993, Novy-Marx 2012)
- Reversal: 1-week, 2-week (Lehmann 1990, Da-Liu-Schaumburg 2014)
- Residual Momentum: 12-week FF-1 residual (Blitz-Huij-Martens 2011)
- Volatility & Skew: IVOL (Ang et al. 2006), MAX (Bali et al. 2011), Realized Volatility
- Defensive & Illiquidity: Betting Against Beta (Frazzini-Pedersen 2014), Cross-Sectional Skewness (Harvey-Siddique 2000), Amihud Illiquidity (Amihud 2002)
The top 7 alphas by |Information Ratio| are selected, sign-corrected, and blended via IC-weighting with 50% shrinkage toward equal weight. Smoothed via EWMA (halflife = 20 days) to minimize week-to-week turnover, the composite alpha yields a robust weekly IR of 0.12.
Seven construction architectures operate on the top-50 most informative tickers per week, leveraging Ledoit-Wolf shrinkage for robust covariance estimation:
- Naïve Equal Weight (Top-Quintile)
- Hierarchical Risk Parity (HRP) - (López de Prado 2016)
- Maximum Diversification - (Choueifaty-Coignard 2008) — Optimal Engine
- Inverse Volatility - (Moreira-Muir 2017 style)
- Robust Alpha Long-Only (CVXPY L2 + Turnover Penalty)
- Robust Alpha Market-Neutral ($\beta=0$, dollar-neutral)
- CVaR Alpha Market-Neutral (Rockafellar-Uryasev 2000)
Statistical validation relies on rigorous tests to eliminate multiple-testing bias and ensure the Sharpe ratio is highly significant and not a product of data snooping.
| Metric | Result | Benchmark (SPY) |
|---|---|---|
| Annualized Sharpe | 2.51 | 0.60 |
| Sortino Ratio | 3.72 | 0.88 |
| Calmar Ratio | 2.60 | - |
| CAGR | 28.4% | 8.7% |
| Annualized Volatility | 12.5% | 15.7% |
| Max Drawdown | -7.7% | -55.2% |
| Test / Metric | Result | Interpretation |
|---|---|---|
| Deflated Sharpe Ratio (DSR) | 0.894 | Corrects for 13 multiple-testing variants; proves statistical significance |
| Probabilistic Sharpe (PSR) | 0.989 | >98.9% probability true Sharpe > 1.0 (Adjusts for non-normality/kurtosis) |
| Stationary Bootstrap (95% CI) | [2.11, 2.92] | 1000-iteration block bootstrap confirms massive robust outperformance |
| Max-Sharpe p-value | < 1% | 0% probability of discovering this Sharpe entirely by chance alone |
Demonstrates exceptional tail-risk mitigation across historically catastrophic market events.
| Market Event | Strategy Return | SPY Return | Strategy MaxDD |
|---|---|---|---|
| Global Financial Crisis (Sep '08 - Mar '09) | -0.3% | -36.9% | -1.1% |
| COVID-19 Crash (Feb-Apr 2020) | -0.3% | -13.2% | -1.6% |
| Rate-Hike Bear Market (H1 2022) | -2.9% | -20.0% | -4.8% |
| 2018 Q4 Selloff | -9.9% | -13.5% | -10.1% |
| 2024 Q3 Volatility Spike | +12.0% | +5.8% | -4.4% |
Tested on unseen market data (with 2024 restricted purely for 252-day indicator warmup) leveraging the exact composite weights and thresholds discovered In-Sample:
- OOS Sharpe Ratio: 2.14
- OOS CAGR: 26.2%
- OOS Max Drawdown: -6.6%
To propel Sharpe from an initial 0.97 to 2.51, two distinct classical scaling mechanisms were incorporated:
- Volatility Targeting: Scales portfolio gross dynamically via
target_vol / realized_vol(capped at 2.5x). Actively de-risks during fat-tail environments while amplifying returns in low-vol regimes. - Dynamic VIX Circuit Breaker:
- VIX ≤ 18 (Normal): Full 1.0x exposure (~83% of days)
- VIX 18–25 (Elevated): Halved 0.5x exposure (~12% of days)
- VIX > 25 (Extreme): 0.0x exposure (Flat to cash) (~5% of days)
(Note: Market-Neutral engines presented negative Sharpe ratios primarily due to short-leg inefficiencies in aggressive trending environments, emphasizing that long-only robust construction paired with strict risk overlays remains structurally superior for this specific universe).
# Clone the repository
git clone https://github.com/yourusername/AlphaStack.git
cd AlphaStack
# Install dependencies (use virtual environment)
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt # Ensure numpy pandas cvxpy yfinance scipy matplotlib are installed
# Execute the Alpha Pipeline
jupyter notebook quant_alpha_search_V3.ipynbDistributed under the MIT License. See LICENSE for more information.