Methods and arrangements to identify feature contributions to erroneous predictions

A system comprising:

• one or more processors;
• memory configured to store instructions, the instructions, when executed by the one or more processors, to cause the one or more processors to:
• determine two or more models from a library of predictive models, each model of the two or more models trained with a training dataset and validated with a testing dataset;
• test the two or more models during a monitoring period, the monitoring period comprising a time frame associated with a monitoring period dataset;
• perform residual modeling on each model in the two or more models during the monitoring period, to determine a list of input features that contribute to a residual for each model of the two or more models;
• rank the list of input features to generate a ranked list of input features;
• generate a feature report identifying each input feature and corresponding classification, the feature report further comprising a classification of each input feature on a scale that ranges from exogenous to endogenous along with probabilities associated with classifications based on counts or weighted counts for the input features;
• and
• display the feature report on a display device including each input feature on the scale that ranges from exogenous to endogenous along with the probabilities associated with the classifications.