Actual versus expected

Actual versus expected (A/E) compares the observed and the predicted totals for each level of each rating factor. It shows where the model and the data disagree: exact balance by factor level is not a general property of boosted models, nor of every GLM.

For each feature, the objects are aggregated over the cells of the grid the deployed tables use (cell 0 holds missing values):

• \(actual = \sum w_i t_i\)
• \(expected = \sum w_i \mu_i\), where \(\mu_i\) includes the exposure for a model with a log link
• \(A/E = actual / expected\)

ae = model.actual_vs_expected(train_data, "ClaimCount", exposure="Exposure")
for factor in ae:
    print(factor["feature"], factor["ae"])

If the model was trained with a sample weight or an exposure, pass them explicitly, even on the training data. See actual_vs_expected.

Pricing report

pricing_report collects, in one document for review, the rating tables, the actual versus expected cells with their axes, and the diagnostics of pruning and graduation:

import json

report = model.pricing_report(test_data, "ClaimCount", exposure="Exposure")
with open("freq_report.json", "w") as f:
    json.dump(report, f)

The report describes the data passed to it; it is not evidence that the data were held out. Save it beside the serialized model.