Model drift is the loss of a model's accuracy or calibration over time because the world it describes has changed, whether in court behavior, procedure or the mix of matters it sees. It is detected by testing on periods after training.
A model is a snapshot of a relationship. When judicial calendars, doctrine, filing patterns or the composition of a book change, the relationship moves and the model's stated likelihoods stop matching what happens. Random train and test splits hide this, which is why chronological splits are used.
Monitoring for drift is routine, not exceptional. Resolved outcomes flow back as new evaluation data, models are retrained on a cadence and re-cleared through the same gate, and a model that has drifted past tolerance is demoted and the registry says so.
Drift is not the same as being wrong on one matter. A well-calibrated model is wrong on individual matters by design. Drift is a population-level change in how well its stated probabilities hold.
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.