What model performance is acceptable?

This depends on the nature and scope of the model. The following criteria can be used to evaluate a satisfactory model performance:

Risk appetite

Number of inputs

Number of hyperparameters

Complexity level

Purpose (e.g. regulatory)

Business model

Here are some specific examples:

Application scorecards

In this case a Gini index of 60% would be good if the portfolio is stable or mature. In case the portfolio is a maturing one, it would be good to have a higher value of Gini as the dynamic in the portfolio can change.

Value at Risk (VaR) back-testing

In case of economic capital, there could be many overshooting. The appropriate amount of 5, 10, 15 or so depends on the materiality of the portfolios for the financial institution.

Neural networks

One of the indicators as the mean absolute percentage error could be 1%, 2% or 5% for example. Again, the acceptable level depends on the effective use of the model and its influence on the strategic decision.