RandomForestRegressor at 0.9 failed live; four books to fix it

The essay is built around a single failure rather than a list of titles. The author had combined blog posts and sample code into a working scikit-learn model, trusted a cross-validation score above 0.9, and shipped it; the first week of operation exposed the gap between the score and actual demand. The diagnosis is that the time-series data had been randomly shuffled before splitting, so future information entered the training side — a validation design error, not a modeling error.

The four books are ordered to match the sequence in which the author hit those problems. The first covers preprocessing, cross-validation and pipelines; the second compresses theory into a short reference; the third, by the Keras author, covers CNN and RNN implementation and what to do when labeled data is scarce; the fourth, by the same author as the second, deals with deployment, monitoring and retraining. The closing observation is that knowledge for squeezing out accuracy and knowledge for building something that does not break are separate things.