# What supervised machine learning model can be used to generate a scorecard-like result?

A scorecard is typically used in Credit Application. One very common model for developing a credit scorecard is logistic regression since it has well-defined probabilities.

Apart from logistic regression, is there any model that can be used in the scorecard?

For example, I don't know whether Support Vector Machine can be used since it only outputs a decision boundary.

More on the scorecard:

• Features are assigned with weightings
• All features are categorical
• The sum of weightings of all features with value True is the total score (like a checklist)
• There will be a cutoff point to classify good/bad (label, +1,-1)
• How far from the cutoff point represents probabilities.
• As you have mentioned that scorecard generation using logistics regression that represent it's more kind of classification task so for it you can use any classification based algorithms like svm, decision tree, random forest and neural network – Swapnil Pote Jun 9 at 19:44
• Other approaches can be used in it is learn to rank approach which generally used inside any search engine or recommendation engine. If you can help me the exact kind output values you are targeting then it would be much better to suggest implementation – Swapnil Pote Jun 9 at 19:49

It depends what you mean by "can be used": any regression algorithm can be used, the question is how reliably it would perform. You can compare different algorithms experimentally (if you have a dataset).

[Updated after question edited]

In general the way to use ML with this kind of setting is to train a classification model based only on the categorical features. Depending on the type of algorithm, the combination of features might not always be a weighted sum, and the result label may or may not be based on a cutoff point. In order to have a cutoff point (thus a numerical prediction), the method must be a soft classification method. Alternatively a regression model could be trained for predicting the numerical value.

So that leaves you with many options:

• soft classification: linear/logistic regression, Naive Bayes, ...
• regression: linear/logistic regression, SVM, decision trees, ...

Note: technically the probability doesn't represent "how far from the cutoff point", it represents the probability of the instance being positive (p=1).

• Maybe I should have added more details on the scorecard. Basically each feature will have a weight and the result (total score) is a binary classification (good/bad). e.g. the score is scaled from 0 to 100. The cutoff point is 50. – JOHN Jun 9 at 0:23