A model that looks good in a notebook can be worthless in trading, and the usual reason is not the algorithm but the evaluation. This skill teaches scikit-learn as a way of working, with short checks (numbers you compute, steps you order, concepts you match) and finance examples: why a scaler must be fitted on the training days only, how overlapping labels leak the future and how purging and an embargo remove it, what ridge and lasso do to coefficients, why a logistic regression for direction barely beats always guessing up, how a forest overfits a weak signal, how to score a return model with the information coefficient, how many fits an honest hyperparameter search needs, and what clustering, PCA and permutation importance can and cannot tell you. The capstone is a code lab where you build the purged walk-forward splitter, the cross-sectional features and the scoring of a monthly stock model.
Comfort with Python, NumPy and pandas is expected, together with basic statistics (mean, variance, correlation, regression).
