Skorecard v1.5.1

Changes

• Corrected the weight plot plotting, I was too fast releasing the previous release :D

Skorecard v1.5.0

Changes

• Added weight plots for coefficients (#65)

Huge thanks to @satya-pattnaik for the contribution!

Skorecard v1.4.0

Changes

• Added random_state to Skorecard, BucketingProcess (#49)

Skorecard v1.3.0

Changes

• Missing value bug fixed: Most and least risky bucket cannot be -1 (#62)

Documentation

• Section on RandomizedGridSearchCV (#26)
• Complete missing value treatment (#25)
• General cleaning

Skorecard v1.2.1

Changes

• Enabled the Skorecard class to also plot weight of evidence (#61), which was not erroneously included in #59

Skorecard v1.2.0

Changes

• Allow plotting of Weight of Evidence in the plot_bucket function #59

Skorecard v1.1.0

Changes

• Add passthrough option for missing_treatment (#57), meaning that missing values can be left alone.
• Validate that passed kwargs are viable (#56)

Skorecard v1.0.0

This release addresses multiple issues, fixes quite some bugs, and overall ensures better scikit-learn compatibility of different transformers and estimators in skorecard. v1.0.0 does include some breaking changes. An overview of the most important changes:

Changes

• BucketingProcess now does not enforce categorical features to be prebucketed, as it is not strictly necessary.
• skorecard.preprocessing.WoEEncoder has now been deprecated in favor of category_encoders.woe.WOEEncoder, which is a new dependency.

These are breaking changes:

• skorecard.Skorecard() dropped support for cat_columns and selected_features. To make sure to bucket categoricals, supply a bucketing estimator. If not supplied, categoricals will be auto-detected. A new variables parameter will select features to be used in the LogisticRegression. A new calculate_stats controls if the LogisticRegression will calculate standards errors or not, and is set to False (previously this was always done).
• skorecard.Skorecard() has an updated default pipeline that will be used if no pipeline is specified. Tweaked some parameters in the bucketing and the detecting of categorical vs numerical columns.
• skorecard.linear_model.LogisticRegression now has a new calculate_stats parameter that control whether or not to calculate (expensive) statistics during .fit(). Defaults to False.
• skorecard.bucketers.DecisionTreeBucketer() has a new parameter dt_kwargs, with keyword arguments passed to the DecisionTreeClassifier. Previously kwargs were directly passed.
• skorecard.bucketers.OptimalBucketer() has a new parameter ob_kwargs with keyword arguments passed to optimalbinning.OptimalBinning. Also several new parameters exposed directly, see docs.
• Input validation on many classes have now been moved to the fit() method, and many internal class variables now have a trailing underscore (_) to signal they are determined during fit. This might cause errors to occur later in your skorecard pipeline. Reason for this is to properly support estimator cloning in scikit-learn, so you can use skorecard estimators in grid searches and cross validations.

Added

• Improved error handling and input validation in many classes
• Added support for numpy matrixes, which are converted gracefully to pd.DataFrame()
• Expanded unit tests to use scikit learns suite of checks in check_estimator

Skorecard v0.7.1

• Allow BucketingProcess() in Skorecard (#51 )

Skorecard v0.7

Added

• Added .fit_interactive() for BucketingProcess (#46)