What Is Feature generation?

Feature generation is the systematic creation of candidate features from raw or lightly processed data. It may use predefined transformations, such as rolling moments and technical indicators, or automated methods that extract a large library of time-series characteristics. Feature generation expands a research search space; it does not establish that a candidate has predictive or trading value.

Generated features must meet the same availability and validation requirements as hand-written features. In particular, candidate generation, filtering, and feature selection must be performed using only the training data of each fold. Because broad generation creates many opportunities for false discoveries, researchers should compare compact feature families against a matched baseline and retain candidates only when they improve out-of-sample results.

See also

• Feature engineering

• Feature selection

• Feature ablation

Research and literature

• Maximilian Christ, Nils Braun, Julius Neuffer, and Andreas W. Kempa-Liehr, Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests, Neurocomputing 307, 2018.