What is elastic net and when is it better than pure LASSO?

Question

AcadiFi · Accepted Answer

Yes — elastic net was designed specifically to address LASSO's behavior with correlated predictors. It blends LASSO's L1 penalty (which does selection) with ridge's L2 penalty (which shrinks smoothly). The objective function is: RSS + λ × [α × Σ|βⱼ| + (1−α) × 0.5 × Σβⱼ²], where α ∈ [0, 1] controls the mix.

α = 1 gives pure LASSO. α = 0 gives pure ridge. α = 0.5 is a common balance. Setting both λ and α requires a 2D cross-validation grid.

Why it helps with correlated predictors: pure LASSO picks ONE variable from a cluster of highly correlated candidates and drives the rest to zero, seemingly at random — the choice can flip with a tiny data change. This makes the model unstable and obscures that the underlying signal comes from a group of related measurements. Elastic net, via the L2 component, encourages the grouping effect: correlated predictors receive similar coefficients rather than arbitrary 0/non-zero splits.

Consider Osselwick Capital modeling credit spreads on 40 predictors including heavily correlated leverage metrics (debt/equity, debt/assets, debt/EBITDA). Pure LASSO picks debt/EBITDA and kills the others. Elastic net at α = 0.4 keeps all three with moderate coefficients, giving a more robust and interpretable picture.

When elastic net wins:
- Many highly correlated predictors
- You want stability across resamples
- You need grouping effect for interpretability

When LASSO is fine:
- Predictors are roughly uncorrelated
- Maximum sparsity is the goal (fewer variables, clearer model)
- Computational speed matters (LASSO is slightly faster)

Both are available in glmnet (R) and scikit-learn (Python) with cross-validation built in.

What is elastic net and when is it better than pure LASSO?

Master Level II with our CFA Course

Related Questions

Practice Questions