Platform For AI:Scorecard training

Feature engineering

The main difference between scorecard training and standard linear modeling is that scorecard training applies feature engineering before fitting the model. The Binning component supports two approaches:

• One-hot encoding: Bins each variable and generates N dummy variables (N = number of bins). You can apply constraints to individual dummy variables.

• WOE conversion: Replaces the original variable value with the weight of evidence (WOE) value of its bin. This encodes predictive information directly into the feature.

Score transformation

In credit scoring, raw model predictions (log-odds) must be converted into interpretable scores. Scorecard Training applies a linear transformation using three parameters:

These three parameters define two points on the score line. For example, with scaledValue=800, odds=50, and pdo=25:

log(50)  = a × 800 + b
log(100) = a × 825 + b

Solve for a and b to derive the linear mapping, then apply it to transform model weights into scores.

Pass the transformation configuration using the -Dscale parameter in JSON format:

{"scaledValue": 800, "odds": 50, "pdo": 25}

All three fields are required when -Dscale is specified.

Constraints

During training, you can add constraints to control how variable weights are learned. Specify constraints in the Binning component — they are automatically passed to Scorecard Training as a JSON string stored in a single-cell table (inputConstraintTableName).

Optimization algorithms

On the Parameters Setting tab, select Advanced Options to configure the optimization algorithm.

If you are unfamiliar with optimization algorithms, set Optimization Method to Auto Selection. The system selects the most appropriate algorithm based on data volume and whether constraints are present.

Feature selection

The component supports stepwise feature selection — a combination of forward and backward selection. After each forward step (adding a variable), a backward pass removes any variables that no longer meet the significance threshold.

Use the following table to choose the right selection standard for your setup:

Marginal contribution measures the difference in objective function value between a model with and without variable X. It applies to all feature engineering and model type combinations, making it the most flexible standard.

Score test (forward): In each iteration, the score chi-square of each candidate variable is computed. The variable with the highest chi-square is added, provided its p-value is below SLENTRY. The backward pass uses Wald chi-square; variables with p-value above SLSTAY are removed.

F test (forward): Similar to score test but uses F distribution. F-values are computed for each candidate variable; variables with p-value above SLENTRY are excluded. The backward pass applies the same F-value logic.

Forced variable selection: Specify variables to include unconditionally — no forward or backward selection is applied to them. Configure this using the -Dselected parameter:

{"max_step": 2, "slentry": 0.0001, "slstay": 0.0001}

If -Dselected is left blank or max_step is set to 0, no feature selection is performed.