Random forest
A learning method that has hundreds of decision trees vote, each built on a randomly drawn sample and subset of variables, stable and not very prone to overfitting.
The random forest, proposed by Leo Breiman in 2001, is a learning method combining a large number of independent decision trees. Each tree is built on a sample drawn at random with replacement from the data, and at each split it considers only a random subset of the available variables. The final prediction is the average of the trees for a continuous value, or the majority vote for a class. This double randomness makes the trees differ from one another, and their average sharply reduces the variance a single tree would have, without increasing bias much. The method holds up well, needs little tuning, tolerates many correlated variables and provides a measure of each variable's importance. It differs from gradient boosting, which builds its trees in sequence to correct errors, whereas the forest builds them in parallel and averages them. In insurance, it often serves as a quick benchmark, an anomaly detector or a variable selection model before building a tariff, and it is widely used in fraud detection and claims triage, where robustness matters more than the last point of performance.
Textbook case: to identify water damage claims likely to become costly, an insurer trains a 500-tree forest on 60 variables. The model flags 8 percent of files at notification, which account for 45 percent of the final cost, and these files are assigned to an adjuster from the start rather than to standard handling.
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.