Feature Transformation and Predictor Selection for Neural Networks

Summary

The article describes ways to transform existing variables and create or select predictors for a neural-network classification task. It discusses logarithms and roots for changing skew, discretization for grouping values, and normalization, alongside outlier removal or capping. Examples compare transformed distributions using skewness, boxplots, missing-value patterns, and relationships among variables. A sine transformation is also presented, with reported sample results showing more symmetrical distributions for the example data.

For predictor selection, the document covers visual and analytical evaluation as well as neural-network pruning. In the worked example, pruning identifies a smaller input set, and the article notes that the selected variables depend on network structure and settings. These demonstrations provide a feature-engineering workflow, not evidence that the resulting predictors improve out-of-sample trading performance. Transformations can alter relationships and require careful treatment of invalid values, missing data, and training versus test data; the examples are specific to the dataset and model used.

Key ideas

• Feature transformations can change scale, distribution shape, and relationships among variables.
• Logarithms can reduce right skew but require care with zero or negative inputs, while roots have different input constraints.
• Outlier removal and capping are alternative treatments whose effects can be compared with distribution and missingness diagnostics.
• Predictor selection can combine visual assessment, analytical methods, and neural-network pruning.
• The selected inputs depend on the model configuration and need validation beyond the demonstrated sample.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.