Ridge Regression

Ridge Regression adds L2 regularization to Ordinary Least Squares (OLS), making it particularly well-suited for A-Share return prediction where multiple factors exhibit high multicollinearity.

---

Overview

The Ridge Regression objective function:

$$\min_{w} |Xw - y|^2 + \alpha |w|^2$$

where $\alpha$ controls regularization strength. When factors are highly correlated (e.g., multiple momentum factors), OLS coefficients diverge. Ridge regression shrinks coefficients through the penalty term, maintaining stable estimation.

Official docs: Ridge Regression — scikit-learn

---

Applications in A-Share Quantitative Strategies

1. Multi-Factor Return Prediction

Feed 50+ Alpha factors as feature matrix $X$ with future 5-day returns as target $y$. Ridge regression remains stable under high factor correlation while OLS coefficients explode.

2. Risk Model Factor Exposure Estimation

In Barra-style risk models, Ridge regression fits individual stock returns against style factor exposures. The $\alpha$ parameter prevents the factor exposure matrix from becoming singular.

3. Automatic Alpha Selection (RidgeCV)

scikit-learn's RidgeCV auto-selects the optimal $\alpha$ via Leave-One-Out cross-validation:

from sklearn.linear_model import RidgeCV
alphas = [0.01, 0.1, 1, 10, 100]
model = RidgeCV(alphas=alphas, cv=5)
model.fit(X_train, y_train)
print(model.alpha_)  # Optimal regularization strength

---

Key Parameters (Finance-Recommended)

Parameter	Description	Recommended
alpha	L2 regularization strength	0.1–100 (log search)
fit_intercept	Fit intercept term	True
solver	Solver algorithm	'auto'

---

Strengths & Limitations

Strengths:

• Closed-form solution — extremely fast training: $w = (X^TX + \alpha I)^{-1}X^Ty$
• Coefficients are directly interpretable as factor weights for strategy attribution
• Robust to multicollinear factors — common in A-Share factor libraries

Limitations:

• Linear assumption cannot capture non-linear factor interactions
• L2 regularization does not produce sparse solutions — all factors are retained
• Sensitive to outliers — apply Winsorization to financial data first

---

Official References

• 📘 scikit-learn Ridge Regression Docs
• 💻 Ridge API
• 📊 RidgeCV API