A challenger model is a candidate trained or configured differently from the model in production and run beside it on the same sealed data, so that a change is made only when the challenger is shown to be better.
Challengers keep a production model honest. They are evaluated on the same holdout under the same registered criteria, and the comparison is recorded. A challenger that wins is promoted through the gate like any other model. One that does not win is kept in the registry with its result.
The practice also guards against change for its own sake. A newer method is not assumed to be better. It has to demonstrate improved discrimination or calibration on data neither model was fit to.
A challenger is not an ensemble and not a second opinion shown to a user. It is a development control that sits behind the production model.
Working a file that turns on this term?
Working through a diligence process?
Institutional partners evaluating a position against this platform's outcome and duration models are welcome to reach out directly.