Fraud risk scoring tool

A method of determining fraud risk scores, the method comprising:

• generating a plurality of customer IDs, each customer ID associated with a customer profile, wherein each customer ID is a randomized, unique identifier associated with transaction metrics of the associated customer profile;
• receiving the transaction metrics indexed with one of the plurality of customer IDs from a network including at least one of an online sales site or an in-store terminal, the transaction metrics including sales data and refund data;
• aggregating the transaction metrics across the plurality of customer IDs to produce an aggregated transaction metric data set;
• comparing each aggregated transaction metric of the aggregated transaction metric data set to historical data from the plurality of customer IDs;
• identifying an outlying transaction in the aggregated transaction metric;
• determining, responsive to the comparison, a precision value for each aggregated transaction metric in the aggregated transaction metric data set, wherein the precision value is an inverse of a variance for the each aggregated transaction metric across the plurality of customer IDs;
• determining, for each aggregated transaction metric in the aggregated transaction metric data set, a suspension propensity importance;
• calculating, for each aggregated transaction metric in the aggregated transaction metric data set, a scoring weight using a trained machine learning (ML) model based at least on one of the precision value, the identified outlying transaction, and the suspension propensity importance, wherein the trained ML model is trained on transaction feedback and includes splitting data into training, validation, and test data sets and training the ML model to maximize a parameter of a precision-recall data set;
• calculating, for each customer ID within the plurality of customer IDs, a risk score based at least on the scoring weight applied to corresponding transaction metrics;
• determining, based at least on the risk score for each customer ID within the plurality of customer IDs, a risk priority;
• reporting, for a selected risk priority, at least one customer ID associated with the selected risk priority to a graphical user interface (GUI) for investigation;
• and
• denying a refund to the at least one customer ID responsive to the reporting of the at least one customer ID.

One or more computer storage devices having computer-executable instructions stored thereon for determining fraud risk scores, which, on execution by a computer, cause the computer to perform operations comprising:

• generating a plurality of customer IDs, each customer ID associated with a customer profile, wherein each customer ID is a randomized, unique identifier associated with transaction metrics of the associated customer profile;
• receiving the transaction metrics indexed with one of the plurality of customer IDs from a network including at least one of an online sales site or an in-store terminal, wherein the transaction metrics comprise sales transaction data, refund transaction data, and cancellation transaction data;
• aggregating the transaction metrics across the plurality of customer IDs to produce an aggregated transaction metric data set;
• comparing each aggregated transaction metric of the aggregated transaction metric data set to historical data from the plurality of customer IDs;
• identifying an outlying transaction in the aggregated transaction metric;
• removing, from the aggregated transaction metric data set, at least a portion of aggregated transaction metrics having a correlation exceeding a correlation threshold;
• determining, responsive to the comparison, a precision value for each aggregated transaction metric in the aggregated transaction metric data set, wherein the precision value is an inverse of a variance for the each aggregated transaction metric across the plurality of customer IDs;
• training a machine learning (ML) model on transaction feedback, wherein training the ML model on transaction feedback comprises:
• splitting data into training, validation, and test data sets;
• and
• training the ML model to maximize a parameter of a precision-recall data set;
• calculating, for each aggregated transaction metric in the aggregated transaction metric data set and using the ML model, a suspension propensity importance;
• calculating, for each aggregated transaction metric in the aggregated transaction metric data set, a scoring weight based at least on one of the precision value, the identified outlying transaction, and the suspension propensity importance;
• determining, for each customer ID within the plurality of customer IDs, a risk score based at least on the scoring weight applied to corresponding transaction metrics;
• determining, based at least on the risk score for each customer ID within the plurality of customer IDs, a risk priority, wherein determining the risk priority comprises assigning each customer ID to one of a predetermined set of risk categories having no more than five categories;
• reporting, for a selected risk priority, at least one customer ID associated with the selected risk priority to a graphical user interface (GUI) for investigation;
• reporting reasons for the selected risk priority associated with the reported at least one customer ID to the GUI;
• and
• denying a refund to the at least one customer ID responsive to the reporting of the at least one customer ID.