convictions about future developments, may do so within the current transparent model structure. The CSO opts to limit itself to estimations based on data from the public domain and not to include extreme scenarios that experts are not agreed upon. 7.3 Alternative parameterisation of the model The model description we give in Appendix A is in line with current actuarial literature. A unique specification of parameters is forced by demanding that the sum of all fitted K’s and kappas equals zero and the sum of all fitted B’s and betas equals one (1). Both choices have evolved historically, although other choices were possible. If one sets not the average of kappas, but the first kappa to zero, this also yields a unique specification of parameters, without the need to average a different number of points at each addition of new data points. That would simplify the comparison of time series over the years. And setting the squared sum of betas to one instead of the sum would eliminate the prerequisite of betas being positive. We would stress that a different choice in these normalisation steps leads to parameter values that describe the same model results in ‘different coordinates’. To preserve compliance with existing literature, we have therefore maintained the notation from the previous publication. 7.4 Consistency over time If the best estimate forecasts from a model that has been estimated deviate little from the actuals observed in subsequent years, one would expect the new parameter estimation to be close to the old parameter values. This desirable property in estimation procedures is sometimes called consistency over time. In reality, even if the best estimate forecasts come true exactly, different values will be found for the new parameters. After all, there are more observations, there are additional assumptions in our model specification (such as the absence of a constant term in the AR process for the Dutch deviation) and we also rescale the K and kappa values to set the average back to zero. Having stated this, the application of the maximum likelihood method implies that we can expect minor adjustments to parameters if new observations are close to earlier projections. It is therefore no surprise that the Projections Life Table AG2018 parameter values are close to the Projections Life Table AG2016 values. Projection Table AG2018 Uncertainty 21

23 Online Touch Home