Feature Engineering

Read these first

1. 01One-Hot Encoding and the High-Cardinality Problem

   One-hot encoding turns a category into a column of 0s and 1s so a model can use it, but it stops working cleanly once a category has hundreds or thousands of possible values, exactly the case for tickers, sectors and counterparties.

2. 02Feature Crosses for Tabular Models

   A feature cross combines two features into one so a model can learn an interaction directly, instead of hoping it discovers on its own that the effect of one variable depends on another.

3. 03Lag and Rolling-Window Features

   Turning a time series into features a model can use means deciding how far back to look and how to summarise what you saw, get the boundary wrong and the model trains on information it wouldn't have had yet.

4. 04Pipelines and Fit/Transform Discipline

   A model that trains beautifully and fails live almost always broke the same rule, some piece of preprocessing was fit on data the model shouldn't have seen yet, or fit differently in training than in production.

5. 05Selecting Features Inside the CV Loop

   Choosing which features to keep before you split into train and test folds leaks information from the test set into the selection decision, the fix is to redo feature selection separately inside every fold.

Then the rest

1. 06Calendar and Datetime Features
2. 07Standardization vs Min-Max Scaling
3. 08Bayesian Smoothing for Rare Categories
4. 09Correlation Filtering and Redundant Features
5. 10Entity-Level Features and Group Leakage
6. 11Feature Binning and Discretization
7. 12Group Aggregation Features
8. 13Informative Missingness and Indicator Features
9. 14Label Construction and Prediction Horizon
10. 15mRMR: Minimum Redundancy Maximum Relevance
11. 16Mutual Information Feature Screening
12. 17Rank and Quantile Transforms
13. 18Recursive Feature Elimination
14. 19Robust Scaling with Median and IQR
15. 20Sample Weights and Recency Decay
16. 21The Hashing Trick
17. 22Transforming the Target Variable
18. 23Unseen Categories at Inference Time
19. 24Weight of Evidence and Information Value