Choosing the RMST horizon
The restriction horizon is part of the prediction question. It should be chosen before fitting when a clinically meaningful decision window is available.
Preferred clinical workflow
Choose τ from the endpoint, intended decision and observed follow-up support. Document the choice before evaluating model performance.
Generic benchmark workflow
When no external horizon is available, the package can derive a quantile from observed event times in the training fold only:
model = SurvFMRMSTRegressor(
backbone="tabpfn",
tau=None,
tau_quantile=0.8,
min_events_for_tau=5,
)
model.fit(X_train, time_train, event_train)
print(model.horizon_selection_.to_dict())
This avoids test-outcome leakage but makes the estimand split-specific. It is a benchmark convenience, not a substitute for a prespecified clinical horizon.