Good features often improve models more than algorithm changes, especially for tabular data.
Date and Time Features
Day of week, month, hour, holidays, time since an event, and cyclical encodings of time for periodic patterns.
Aggregates
Summaries over related records: a customer's average order value, number of purchases in 30 days, days since last login.
Ratios and Differences
Price per square metre, spend relative to average, change since last month.
Interactions
Combinations of features, such as region and product category together, when effects depend on both.
Binning
Grouping continuous values into ranges, useful for some models and for interpretability.
Text Features
Word counts, keyword flags, TF-IDF vectors or embeddings from language models.
Domain Knowledge
Experts often know which signals matter. Talk to them.
Avoid Leakage
Features must use only information available at prediction time. Compute aggregates using data from before the prediction date.
Validate
Measure each feature's contribution with cross-validation; remove features that add complexity without value.