There are two broad types of risk scores:
- the sum scores and
- the scores estimated by statistical models.
|Sum scores||These risk scores are the sum of the individual weights given to the scoring rules hit by the transaction. The sum score adds points to a total. If the total of the sum score is higher than a threshold, the order is sent for verification; the treatment in case of a high score is similar to forcing rules.|
|Statistical models||These risk scores are statistical models optimized numerically. The outcome of statistical models to predict risk is either a number between 0 (0%) and 1 (100%), which reflects the likelihood that a payment is a fraud or a class, e.g., fraud/not-fraud.|
In terms of pros and cons, sum scores are simple to apprehend and inexpensive to set up. Frequently, this is the first way to build a risk score. However, that simplicity leads to a poor calibration.
Hence, because the weights of the sum scores are set manually, they are poorly calibrated with payment fraud. On the other hand, estimating the weights of the statistical model with optimization algorithms can lead to risk scores whose calibrations improve dramatically compared to that of a sum score.
However, the better calibration of statistical models comes at the expense of complexity and expertise because statistical models are much harder to understand, set up, and manage, especially if they are non-linear. Still, the boost in performance has the potential to greatly offset the cost, and for linear statistical models, people with expertise in statistics will know how to interpret them.