Two-stage regression for non-negative targets with a point mass at zero.predict defaults to the mixture median, suitable for absolute error. With
positive probability p, it is zero for p <= 0.5 and otherwise the positive
distribution’s (p - 0.5) / p quantile. Quantiles use linear interpolation on
the supplied grid, clamping levels outside that grid. Negative
positive-stage predictions are clipped to zero. Mean predictions multiply
the clipped positive-stage mean by p.Fitted attributes include hurdle_, zero_rate_, classifier_, and regressor_.
classifier_ is None when the hurdle is inactive. Both estimators are None
for an all-zero target with the hurdle enabled.
Cloneable classifier with predict_proba. Defaults to TabPFN.
regressor
Any
None
Cloneable regressor with TabPFN’s output_type and quantiles prediction arguments. Defaults to TabPFN.
hurdle
Literal[“auto”] | bool
"auto"
Whether to split zero and positive targets. “auto” enables the split for non-negative targets whose zero rate exceeds zero_threshold. Otherwise, all rows go to the regressor.
zero_threshold
float
0.25
Training zero-rate threshold for automatic selection. Defaults to 0.25, a heuristic rather than a tuned optimum. This controls whether to fit two stages, independently of the prediction statistic.
Strictly increasing positive-stage quantile levels between 0 and 1, with at least two entries. None uses the 21-point grid from 0.01 to 0.99. Required levels outside the grid clamp to its endpoints.