OptiAlpha: Convex Regularized Alpha Combination and Dynamic Multi-Factor Allocation Engine

OptiAlpha is an institutional-grade Python framework for regularized multi-factor alpha combination, dynamic covariance shrinkage, and online signal allocation. It solves the multi-factor aggregation dilemma by trading off expected predictive Information Coefficients (IC), factor portfolio risk, $L_1$ turnover transaction costs, and factor concentration constraints via accelerated proximal gradient methods (FISTA).

---

The Quantitative Pipeline: OrthoQuant $\to$ OptiAlpha

In modern quantitative equity and multi-asset factor investing, the signal construction pipeline follows a rigorous two-stage architecture:

[ Raw Alpha Factors ]
         │
         ▼
┌─────────────────────────────────────────────────────────────┐
│ 1. OrthoQuant (Signal Pre-Processing & Clean-up)            │
│    • Cross-Sectional Risk Neutralization (FWL via Thin QR)  │
│    • Lowdin Symmetric Orthogonalization (F^T F = I_K)       │
└─────────────────────────────────────────────────────────────┘
         │
         ▼  (Orthonormal Residual Factor Streams)
┌─────────────────────────────────────────────────────────────┐
│ 2. OptiAlpha (Alpha Combination & Factor Allocation)        │
│    • Ledoit-Wolf Analytical Covariance Shrinkage            │
│    • Convex L1-Regularized Turnover Optimization (FISTA)    │
│    • Online Exponentiated Gradient Adaptive Weighting       │
│    • Meucci Effective Number of Alphas (ENF) Attribution    │
└─────────────────────────────────────────────────────────────┘
         │
         ▼
[ Optimized Composite Alpha Score w_t^T F_t ]

---

Mathematical Notation Reference

Symbol	Type	Description
$N \in \mathbb{N}$	Dimension	Number of assets in the universe ($N \ge K$)
$K \in \mathbb{N}$	Dimension	Number of candidate alpha factors
$T \in \mathbb{N}$	Dimension	Number of historical observation periods (trading days)
$\mathbf{F}_t \in \mathbb{R}^{N \times K}$	Matrix	Cross-sectional factor exposure matrix at time $t$
$\mathbf{ic} \in \mathbb{R}^K$	Vector	Information Coefficient vector ($\mathrm{Rank\ IC}$)
$\mathbf{\Omega}_{\alpha} \in \mathbb{R}^{K \times K}$	Matrix	Factor covariance matrix
$\mathbf{w} \in \mathbb{R}^K$	Vector	Factor allocation weight vector
$\mathbf{w}_{\text{prev}} \in \mathbb{R}^K$	Vector	Previous rebalance weights for turnover control
$\gamma \ge 0$	Scalar	Risk aversion penalty on factor portfolio variance
$\lambda_{\text{turnover}} \ge 0$	Scalar	$L_1$ transaction cost penalty
$\lambda_2 \ge 0$	Scalar	$L_2$ ridge regularization parameter
$\lVert \cdot \rVert_1, \lVert \cdot \rVert_2$	Norms	Vector $L_1$ and $L_2$ norms
$\lVert \cdot \rVert_F$	Norm	Matrix Frobenius norm

---

Mathematical Architecture & Theoretical Formulations

1. Fundamental Law of Active Management & Optimal Alpha Allocation

Let $\mathbf{F}_t \in \mathbb{R}^{N \times K}$ be the cross-sectional factor exposure matrix for $N$ assets across $K$ candidate alpha factors at time $t$.
Let $\mathbf{r}_{t+1} \in \mathbb{R}^N$ be the forward asset returns.
Let $\mathbf{ic} = (\mathrm{IC}_1, \dots, \mathrm{IC}_K)^T \in \mathbb{R}^K$ denote the vector of forecast Information Coefficients ($\mathrm{Rank\ IC}$).
Let $\mathbf{\Omega}_{\alpha} \in \mathbb{R}^{K \times K}$ be the factor covariance matrix.

According to the Fundamental Law of Active Management (Grinold & Kahn, 2000; Qian & Hua, 2004), the unconstrained weight vector $\mathbf{w}^*$ that maximizes the composite Information Ratio (IR) satisfies:

$$ \mathbf{w}^* = \mathbf{\Omega}_{\alpha}^{-1} \mathbf{ic} \qquad \text{(1)} $$

The maximal achievable composite Information Ratio is:

$$ \mathrm{IR}^* = \sqrt{\mathbf{ic}^T \mathbf{\Omega}_{\alpha}^{-1} \mathbf{ic}} \qquad \text{(2)} $$

Theorem 1 (The OrthoQuant Closed-Form Connection)

If factors $\mathbf{\widetilde{F}}$ are orthonormalized via OrthoQuant's Lowdin Symmetric Orthogonalization ($\mathbf{\widetilde{F}}^T \mathbf{\widetilde{F}} = \mathbf{I}_K$), then the factor correlation matrix collapses to the identity matrix $\mathbf{C} = \mathbf{I}_K$. Consequently:

$$ \mathbf{\Omega}_{\alpha} = \mathrm{diag}(\sigma_1^2, \dots, \sigma_K^2) \implies w_k^* = \frac{\mathrm{IC}_k}{\sigma_k^2} \qquad \text{(3)} $$

Proof: Substituting $\mathbf{\Omega}_{\alpha} = \mathbf{D}_{\sigma^2}$ into Equation (1), $\mathbf{\Omega}_{\alpha}^{-1} = \mathbf{D}_{1/\sigma^2}$, yielding $w_k^* = \mathrm{IC}_k / \sigma_k^2$. Multicollinearity and ill-conditioned matrix inversion are completely eliminated. $\blacksquare$

---

2. Convex Regularized Objective with Turnover & Concentration Bounds

In practical portfolio management, factor allocations cannot change drastically from period to period due to market friction, slippage, and turnover costs. Furthermore, single-factor concentration must be bounded.

OptiAlpha formulates factor combination as a strictly convex optimization problem:

$$ \min_{\mathbf{w} \in \mathcal{W}} \quad \mathcal{L}(\mathbf{w}) = \underbrace{-\mathbf{ic}^T \mathbf{w}}_{\text{Predictive Return}} + \underbrace{\frac{\gamma}{2} \mathbf{w}^T \mathbf{\Omega}_{\alpha} \mathbf{w}}_{\text{Factor Risk Penalty}} + \underbrace{\lambda_{\text{turnover}} \lVert \mathbf{w} - \mathbf{w}_{\text{prev}} \rVert_1}_{\text{L1 Turnover Friction}} + \underbrace{\frac{\lambda_2}{2} \lVert \mathbf{w} \rVert_2^2}_{\text{L2 Ridge Shrinkage}} \qquad \text{(4)} $$

subject to convex budget and box constraints:

$$ \mathcal{W} = \lbrace \mathbf{w} \in \mathbb{R}^K : l_k \le w_k \le u_k ; \forall k, \quad \sum_{k=1}^K w_k = s \rbrace \qquad \text{(5)} $$

where:

• $\gamma \ge 0$: Risk aversion parameter penalizing factor portfolio volatility.
• $\lambda_{\text{turnover}} \ge 0$: $L_1$ turnover penalty against prior allocation $\mathbf{w}_{\text{prev}}$.
• $l_k, u_k$: Lower/upper bounds ($l_k = 0, u_k = 1$ enforces a long-only simplex $\Delta^K$).

---

3. Accelerated Proximal Gradient Solver (FISTA)

Because the $L_1$ turnover term $\lVert \mathbf{w} - \mathbf{w}_{\text{prev}} \rVert_1$ is non-smooth, classical gradient descent fails. OptiAlpha implements the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) (Beck & Teboulle, 2009) with $\mathcal{O}(1/k^2)$ optimal convergence rate:

Step 1: Smooth Gradient Step

$$ \nabla f(\mathbf{y}^{(k)}) = -\mathbf{ic} + \gamma \mathbf{\Omega}_{\alpha} \mathbf{y}^{(k)} + \lambda_2 \mathbf{y}^{(k)}, \quad \mathbf{z}^{(k)} = \mathbf{y}^{(k)} - \frac{1}{L} \nabla f(\mathbf{y}^{(k)}) $$

where $L = \lVert \gamma \mathbf{\Omega}_{\alpha} + \lambda_2 \mathbf{I} \rVert_2$ is the exact Lipschitz constant computed via spectral norm.

Step 2: Shifted Soft-Thresholding ($L_1$ Proximal Operator)

$$ \mathbf{v}^{(k)} = \mathbf{w}_{\text{prev}} + \mathrm{sign}(\mathbf{z}^{(k)} - \mathbf{w}_{\text{prev}}) \odot \max \left( \lvert \mathbf{z}^{(k)} - \mathbf{w}_{\text{prev}} \rvert - \frac{\lambda_{\text{turnover}}}{L}, , 0 \right) $$

Step 3: Box-Constrained Simplex Projection

$$ \mathbf{w}^{(k+1)} = \Pi_{\mathcal{W}}(\mathbf{v}^{(k)}) = \mathrm{clip}(\mathbf{v}^{(k)} - \mu^*, \mathbf{l}, \mathbf{u}) $$

where dual parameter $\mu^*$ is solved via bisection root finding such that $\sum_{i=1}^K w_i = s$.

Step 4: Nesterov Momentum Acceleration

$$ t_{k+1} = \frac{1 + \sqrt{1 + 4 t_k^2}}{2}, \quad \mathbf{y}^{(k+1)} = \mathbf{w}^{(k+1)} + \left( \frac{t_k - 1}{t_{k+1}} \right) (\mathbf{w}^{(k+1)} - \mathbf{w}^{(k)}) $$

---

4. Analytical Ledoit-Wolf Covariance Shrinkage

Sample covariance matrices $\mathbf{S} = \frac{1}{T-1} \mathbf{X}^T \mathbf{X}$ are notoriously ill-conditioned when $T$ is small relative to factor count $K$. OptiAlpha computes the asymptotically optimal shrinkage intensity $\delta^* \in [0, 1]$ (Ledoit & Wolf, 2004):

$$ \mathbf{\Sigma}^* = \delta^* \mathbf{T} + (1 - \delta^*) \mathbf{S} \qquad \text{(6)} $$

• Constant Correlation Target ($\mathbf{T}$): Preserves individual sample variances $S_{ii}$ while shrinking cross-sectional correlations to the mean correlation $\bar{r} = \frac{2}{K(K-1)} \sum_{i < j} R_{ij}$.
• Optimal Shrinkage Parameter:

$$ \delta^* = \min \left( 1, , \max \left( 0, , \frac{\hat{\pi} - \hat{\rho}}{T \cdot \lVert \mathbf{S} - \mathbf{T} \rVert_F^2} \right) \right) $$

---

5. Online Adaptive Allocation (Exponentiated Gradient)

For real-time streaming environments, OptiAlpha provides an online multiplicative update rule (Freund & Schapire, 1997; Kivinen & Warmuth, 1997) guaranteeing sublinear regret $\mathcal{O}(\sqrt{T \ln K})$:

$$ w_{k, t+1} = (1 - \varepsilon) \frac{w_{k, t} \exp(\eta \cdot \hat{r}_{k, t})}{\sum_{j=1}^K w_{j, t} \exp(\eta \cdot \hat{r}_{j, t})} + \frac{\varepsilon}{K} \qquad \text{(7)} $$

where $\eta > 0$ is the learning rate and $\varepsilon \in [0, 1]$ prevents premature factor starvation.

---

6. Risk Attribution & Diversification Metrics

Euler's Risk Decomposition

Total factor portfolio volatility $\sigma_p = \sqrt{\mathbf{w}^T \mathbf{\Omega}_{\alpha} \mathbf{w}}$ decomposes additively into Percentage Risk Contributions ($\mathrm{PRC}_k$):

$$ \mathrm{PRC}_k = \frac{w_k (\mathbf{\Omega}_{\alpha} \mathbf{w})_k}{\mathbf{w}^T \mathbf{\Omega}_{\alpha} \mathbf{w}}, \quad \sum_{k=1}^K \mathrm{PRC}_k = 1.0 \qquad \text{(8)} $$

Effective Number of Alphas (ENF)

Following Meucci (2009), the effective diversification entropy across alpha streams is:

$$ \mathrm{ENF} \triangleq \exp \left( -\sum_{k=1}^K p_k \ln p_k \right) \in [1, K] \qquad \text{(9)} $$

where $p_k = \max(\mathrm{PRC}_k, 0)$.

---

Strategy Comparison Matrix

Strategy	Objective Function	Turnover Control	Handles Multicollinearity	Optimal Solver
Equal Weighting (1/K)	Heuristic benchmark	None	Poor	Closed-form $\mathcal{O}(1)$
IC-Weighted	$w_k \propto \max(0, \mathrm{IC}_k)$	None	Moderate	Closed-form $\mathcal{O}(K)$
Grinold-Kahn Analytical	$\mathbf{\Omega}^{-1} \mathbf{ic}$	None	Requires Invertible $\mathbf{\Omega}$	Linear Solve $\mathcal{O}(K^3)$
OptiAlpha Convex (FISTA)	$\min -\mathbf{ic}^T \mathbf{w} + \frac{\gamma}{2} \mathbf{w}^T \mathbf{\Omega} \mathbf{w} + \lambda \lVert \Delta \mathbf{w} \rVert_1$	Rigorous $L_1$ Penalty	Excellent (L2 Ridge + Ledoit-Wolf)	FISTA $\mathcal{O}(K \log K)$ per iter
OptiAlpha Online (EG)	Multiplicative Weights Online Regret	Adaptive Decay	Good	Online $\mathcal{O}(K)$ per step

---

Installation

git clone https://github.com/huuan26/OptiAlpha.git
cd OptiAlpha
pip install -e .

---

Quickstart

import numpy as np
import pandas as pd
from optialpha import (
    generate_synthetic_alpha_panel,
    ledoit_wolf_shrinkage,
    convex_alpha_combination,
    equal_weight_combination,
    information_coefficient,
    effective_number_of_alphas,
    turnover_ratio,
)

# 1. Generate multi-factor panel (100 assets, 5 factors, 120 periods)
panel = generate_synthetic_alpha_panel(n_assets=100, n_factors=5, n_periods=120)
factors_t0 = panel["factors_panel"][0]
returns_t0 = panel["returns_panel"][0]
ic_history = panel["ic_history"]

# 2. Estimate cleaned covariance matrix via Ledoit-Wolf Shrinkage
shrunk_cov, shrinkage_delta = ledoit_wolf_shrinkage(ic_history, target="constant_correlation")
print(f"Optimal Ledoit-Wolf Shrinkage Intensity (delta): {shrinkage_delta:.4f}")

# 3. Forecast Information Coefficients (Rolling Mean IC)
forecast_ic = ic_history.mean(axis=0).to_numpy()

# 4. Solve Regularized Convex Alpha Combination with Turnover Penalty
w_prior = np.array([0.2, 0.2, 0.2, 0.2, 0.2])
res = convex_alpha_combination(
    factors=factors_t0,
    ic_vector=forecast_ic,
    covariance=shrunk_cov,
    risk_aversion=1.5,
    turnover_penalty=0.03,
    w_prev=w_prior,
    lower_bound=0.0,
    upper_bound=0.40,
    solver="fista",
)

print("\n--- OptiAlpha Allocation Results ---")
print("Optimal Factor Weights:", np.round(res.weights, 4))
print(f"Convergence: {res.converged} in {res.iterations} iterations")
print(f"Predicted Information Coefficient: {res.predicted_ic:.4f}")
print(f"Realized Turnover from Prior Allocation: {res.turnover:.4f}")

# 5. Evaluate Composite Alpha vs Equal-Weight Benchmark
composite_ic = information_coefficient(res.composite_alpha, returns_t0, method="spearman")
eq_alpha, _ = equal_weight_combination(factors_t0)
eq_ic = information_coefficient(eq_alpha, returns_t0, method="spearman")

print(f"\nEqual-Weight Rank IC: {eq_ic:.4f}")
print(f"OptiAlpha Composite Rank IC: {composite_ic:.4f}")

# 6. Measure Effective Diversification (Meucci ENF)
enf = effective_number_of_alphas(res.weights, shrunk_cov)
print(f"Effective Number of Alphas (ENF): {enf:.2f} / 5.00")

---

Testing & Verification

Run the full test suite with coverage report:

pytest

---

Academic References & Bibliography

1. Grinold, R. C., & Kahn, R. N. (2000). Active Portfolio Management: A Quantitative Approach for Producing Superior Returns and Controlling Risk (2nd ed.). McGraw-Hill.
2. Qian, E. E., & Hua, R. H. (2004). Active risk and information ratio in multi-factor models. The Journal of Portfolio Management, 30(3), 20–34.
3. Ledoit, O., & Wolf, M. (2004). A well-conditioned estimator for large-dimensional covariance matrices. Journal of Multivariate Analysis, 88(2), 365–411.
4. Beck, A., & Teboulle, M. (2009). A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2(1), 183–202.
5. Duchi, J., Shalev-Shwartz, S., Singer, Y., & Chandra, T. (2008). Efficient projections onto the $l_1$-ball for predictors in high dimensions. ICML, 272–279.
6. Meucci, A. (2009). Managing diversification. Risk, 22(5), 74–79.
7. Choueifaty, Y., & Coignard, Y. (2008). Toward maximum diversification. The Journal of Portfolio Management, 35(1), 40–51.
8. Freund, Y., & Schapire, R. E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1), 119–139.

---

Author & License

• Author: Phan Bùi Hữu Ân (huuan26)
• License: MIT License - see the LICENSE file for details.