Unsupervised learning
A family of methods that look for structure in data without a target variable, clusters, principal dimensions, anomalies, useful for segmenting a portfolio, building a territorial rating scheme or spotting atypical behavior.
Supervised learning learns to predict a known variable, claim or not, cost, fraud, from labeled examples. Unsupervised learning works without a target: it looks for structure in the data themselves. Its main families are clustering, which forms groups of similar observations, for example with k-means or hierarchical clustering; dimension reduction, which summarizes many correlated variables into a few axes, such as principal component analysis; and anomaly detection. In insurance these methods serve tasks where one does not yet know what one is looking for. They group municipalities or grid cells with similar characteristics to build a smoothed territorial rating scheme, segment customers by behavior, identify families of claims with similar profiles, or summarize hundreds of external variables before building them into a tariff. Their main limit is interpretation: the groups obtained depend on the variables and parameters chosen, have no objective truth, and must be validated by business sense before informing a decision.
Textbook case: to overhaul its motor territorial rating scheme, an insurer describes each of 35,000 municipalities with 25 variables, density, traffic, crime, smoothed claims, then groups them into 12 homogeneous classes by clustering. Actuaries then check that each class makes sense and has distinct claims experience before building it into the tariff.
unsupervised learning, clustering
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.