Speaker(s): Frank van Berkum (UvA)
Pension funds must adjust population-wide mortality forecasts for the valuation of their liabilities, since the combined effect of the many factors that determine their participants' survival probabilities typically differs from the average in the population. Candidates to explain this difference include age, salary and education, but also spatial effects based on longitude and latitude coordinates. We identify the most important factors to improve the accuracy of predicted survival in pension funds, by applying Poisson regressions in a Generalized Additive Model.
For a Dutch pension portfolio with over 11 million observations from individuals, we find that the most relevant risk factors are salary, disability and an allowance for working at irregular hours. It turns out that there is a substantial difference between the expected remaining life expectancies in the most-favorable and least-favorable risk profiles based on this information: about 12 years at age 25 and 10 years at age 65.
We also formulate a financial backtest to determine which risk factors contribute most to the prediction accuracy of the liabilities, which may be more relevant in actuarial applications than the prediction accuracy of individual deaths. In line with expectations, salary information is critical to ensure the liabilities are at the right level, but for specific risk profiles the predictions may be far off. When we include other risk factors such as disability information, liabilities are more accurately predicted for the different risk profiles.