Is Your Model Time-Blind? The Case for Cyclical Feature Encoding

Standard numerical encoding of cyclical features, such as the hour of the day, creates an artificial 'midnight paradox' where values like 23:59 and 00:01 appear far apart to a machine learning model. This issue can be solved by treating time as a circle instead of a line, using trigonometric functions for feature encoding. By transforming a single time variable into two new features using sine and cosine, each point in the cycle is mapped to a unique coordinate on a circle, preserving the natural proximity of boundary values. A code example using a Random Forest Regressor on an energy prediction dataset demonstrates an improvement in model performance by applying this technique. This method is beneficial for various model types and can be used for any cyclical data, including day of the week or month of the year.