Our postcode model is calibrated to the combined mortality data of over 200 defined benefit pension plans, with wide coverage across socio-economic types, industries and geographic regions in the UK.
Using ‘leave-one-out cross-validation’, we test that our calibration of the postcode model works in practice, i.e. that it actually predicts pension scheme mortality as accurately as possible.
The average discrepancy in this test across all pension schemes in the databank gives a fair, objective measure of the model’s actual predictive power for new schemes outside the calibration dataset.
Cross-validation is critical because it:
- tests the model’s purpose, i.e. prediction of mortality at pension plan level – in contrast, in-sample tests are notorious for overstating predictive power, and
- allows us to compare the performance of alternative models (e.g. models based on different rating factors) – cross validation is how we select our model.
Finally, cross-validation also provides an objective measure of how predictive the model is. This means that we can place a confidence interval around the mortality rates produced by the model, which
- is useful for assessing the uncertainty of the results, and
- enables the user to combine the postcode model with the plan-specific mortality experience e.g. by weighting the results using statistical credibility.
Our postcode analysis produces a robust best-estimate mortality basis, with a transparent measure of uncertainty based on leave-one-out cross-validation.