How does one-class SVM work for outlier detection?

Question

AcadiFi · Accepted Answer

One-class SVM learns a boundary around the region containing normal data in feature space, labeling points outside as outliers. Unlike standard SVMs, it has no negative class; it fits a decision function that is positive on the bulk of training data and negative on rare points. The nu parameter (between 0 and 1) upper-bounds the fraction of training points treated as outliers and lower-bounds the fraction of support vectors.

The algorithm uses a kernel (typically RBF) to map inputs into a high-dimensional space and finds a hyperplane that separates the origin from the data with maximum margin. Equivalently, it can be formulated as finding the smallest-volume hypersphere enclosing the normal points.

For daily risk monitoring at a firm called Oathridge Risk, features per trading day might include realized vol, skew, volume imbalance, and credit spread moves. Train one-class SVM on 1,200 days of normal regime data with nu equal to 0.05 and RBF gamma tuned via cross-validation. The model then scores new days: a negative score indicates an unusual day that warrants escalation.

One-class SVM works best when the normal region is compact and well-defined. It is more computationally expensive than isolation forest (O(n squared) to O(n cubed) in training) and highly sensitive to kernel and nu choice. It struggles on high-dimensional sparse data. For CFA candidates, one-class SVM is a solid choice when training labels are entirely positive, the dataset is moderate in size (under 20,000 points), and a crisp decision boundary is desirable. For larger datasets prefer isolation forest; for contextual outliers prefer LOF.

How does one-class SVM work for outlier detection?

Master Level II with our CFA Course

Related Questions

Practice Questions