Exposure and offsets

Exposure

For rate models, such as claim frequency, the objects are observed over different exposures (for example, policy durations). Pass the exposure to fit instead of dividing the target by it:

model = TBoostRegressor(objective="poisson")
model.fit(train_data, "ClaimCount", exposure="Exposure")

The model then predicts a rate per unit of exposure: the predicted rate is multiplied by the exposure before it is compared with the target, which adds \(\log(exposure)\) to the raw score. The exposure is supported for the poisson, gamma and tweedie objectives and for binary classification.

predict returns the rate per unit of exposure, and an exposure column in the data is ignored when predicting. The expected total for an object is predict(X) * exposure. Pass the exposure to predict_contributions to decompose the expected total instead of the rate.

The exposure also weights the data the tables are purified on (see Rating tables), and the support of every table cell is the exposure-weighted count of its objects.

Offset

An offset is a link-scale value added to the raw score of each object and never learned, the equivalent of XGBoost's base_margin and LightGBM's init_score. Use it, for example, to train a model on top of an existing rating structure:

model.fit(train_data, "ClaimCount", exposure="Exposure", offset="BaseLogRate")
preds = model.predict(test_data, offset="BaseLogRate")

Prediction methods add the offset only when it is passed to them again; otherwise objects are scored with a zero offset. For the log and logit links, the offset is added to the exposure offset.

Sample weights

sample_weight multiplies each object's contribution to the loss. For rate models, use the exposure rather than a weight to express the observation period: the exposure enters the prediction, a weight does not.