Box Cox
Box-Cox Transformation is a power transformation that automatically finds the optimal mathematical function to normalize data — searching over a parameter λ (lambda) to determine whether the data needs a log transform (λ=0), square root (λ=0.5), reciprocal (λ=-1), no transform (λ=1), or any other power between them, making it the data-driven alternative to manually guessing which transformation to apply to skewed features.
What Is the Box-Cox Transformation?
- Definition: A family of power transformations parameterized by λ that transforms the data to be as close to a normal distribution as possible — the algorithm finds the optimal λ using maximum likelihood estimation.
- The Formula:
- If $\lambda
eq 0$: $y_{new} = \frac{y^{\lambda} - 1}{\lambda}$
- If $\lambda = 0$: $y_{new} = \log(y)$
- Why Not Just Use Log?: Log transformation assumes the data needs logarithmic compression. But some data needs square root (λ=0.5), cube root (λ=0.33), or even no transformation (λ=1). Box-Cox finds the optimal power automatically.
Lambda Values Explained
| λ (Lambda) | Transformation | When Optimal | Effect |
|---|---|---|---|
| -1 | Reciprocal ($1/y$) | Heavily right-skewed | Extreme compression |
| -0.5 | Reciprocal square root ($1/\sqrt{y}$) | Very right-skewed | Strong compression |
| 0 | Log($y$) | Moderately right-skewed | Logarithmic compression |
| 0.5 | Square root ($\sqrt{y}$) | Mildly right-skewed | Mild compression |
| 1 | No transformation ($y$ itself) | Already normal | No change needed |
| 2 | Square ($y^2$) | Left-skewed | Expansion (rare) |
How Box-Cox Finds Optimal λ
| Step | Process |
|---|---|
| 1. Try many λ values | Test λ from -5 to +5 in small increments |
| 2. For each λ, transform data | Apply $y^{(\lambda)}$ formula |
| 3. Measure normality | Log-likelihood of the transformed data under a normal distribution |
| 4. Select best λ | The λ that maximizes log-likelihood (makes data most normal) |
Python Implementation
from scipy.stats import boxcox
from sklearn.preprocessing import PowerTransformer
# SciPy (returns transformed data + optimal lambda)
data_transformed, optimal_lambda = boxcox(data)
print(f"Optimal lambda: {optimal_lambda:.2f}")
# Scikit-learn (fits inside pipeline, handles inverse)
pt = PowerTransformer(method='box-cox') # requires positive data
X_transformed = pt.fit_transform(X)Box-Cox vs Yeo-Johnson
| Property | Box-Cox | Yeo-Johnson |
|---|---|---|
| Input requirement | Strictly positive ($y > 0$) | Any value (positive, zero, negative) |
| Zero handling | Cannot handle zeros | Yes |
| Negative values | Cannot handle | Yes |
| Optimal for | Positive continuous data | General-purpose |
| Scikit-learn | `PowerTransformer(method='box-cox')` | `PowerTransformer(method='yeo-johnson')` |
When to Use
| Use Box-Cox / Yeo-Johnson | Don't Use |
|---|---|
| Linear models that assume normality | Tree-based models (don't need normality) |
| Right or left-skewed features | Already normally distributed data |
| When you don't know which transform to apply | When you know log transform is correct |
| Preprocessing for statistical tests | Categorical or binary features |
Box-Cox Transformation is the automated alternative to manual transformation selection — finding the optimal power parameter λ through maximum likelihood estimation to produce the most normal-like distribution possible, with Yeo-Johnson as its generalization that handles the zero and negative values that Box-Cox cannot.