Writing your own estimator

The estimator contract

Every scikit-learn model follows one contract, and that contract is why models can be swapped, chained and tuned. __init__ only stores the parameters it receives, under the same names, with no checks and no computation: that is what lets get_params, set_params and clone work. fit(X, y) learns from the data, stores what it learned in attributes that end with an underscore (coef_, intercept_, n_features_in_) and returns self, so calls can be chained. predict(X) first calls check_is_fitted, which raises NotFittedError if fit has not run, then validates X with check_array and applies what was learned. Inherit from BaseEstimator for the parameter handling and from RegressorMixin for a free score method (put the mixin first).

The model in this lab is ridge regression. With centred data it solves (XᵀX+αI) w=Xᵀy, and the intercept is ȳ-x̄·w. A larger α shrinks the coefficients towards zero, which is a good habit when the features are noisy lagged returns. Because your class follows the contract, make_pipeline(StandardScaler(), ShrinkageRegressor()), cross_val_score and GridSearchCV accept it, and clone(model) gives a fresh unfitted copy. That last point matters in a walk-forward evaluation: fit a clone on each training window, never the model itself, or the information of one fold leaks into the next. The score for a return forecast is the Spearman rank correlation between predictions and outcomes on the test rows, the information coefficient.

• Parameters in, learned attributes out: __init__ stores, fit learns into names ending with an underscore and returns self.
• Check before you predict: check_is_fitted and check_array give the standard errors that pipelines and users expect.
• Fit a clone: in a walk-forward loop, clone(model).fit(...) keeps folds independent.

Lab · what's graded

• Lagged features10 pts
• Ridge from the normal equations15 pts
• The constructor10 pts
• Fit and predict20 pts
• Working with Pipeline and GridSearchCV10 pts
• A walk-forward rank correlation15 pts
• Does the past of SPY predict its next day?20 pts