Expectations v. extrapolations
The CMI has published two working papers along with its new mortality projection model. The proposed new model blends current improvement rates into an expected long-term average rate of improvement. The hope is that such models can incorporate expert opinions on mortality trends to improve the accuracy of projections.
This sort of model is described as an expectation by Booth and Tickle (2008), who distinguish it from the commoner approach of extrapolating recent trends. As with cause-of-death-based forecasting, expectation has an intuitive appeal which fades upon closer examination:
Expectation is not generally a good basis for mortality forecasting, as it is subjective; expert expectations are invariably conservative [...] The disadvantage is its subjectivity and potential for bias.
Booth and Tickle (2008), Annals of Actuarial Science 3, I/II 3-43
Booth and Tickle (2008) mention an interesting feature of expectations called "assumption drag", i.e. where expert expectations often prove to lag actual experience, rather than lead it as might be hoped for. Another problem is the phenomenon of "expert flocking", which arises from a common information base and individuals' reluctance to stray too far from consensus.
The choice of an expectation model is a departure from the CMI's more recent tendency to use stochastic projection models, as outlined in other working papers. Indeed, the CMI itself had a recent example where an expectation model failed and had to be replaced: the CMIR17 projections in 1999 were also an expectation based on convergence to long-term improvement rates, but they were quickly superceded in 2002 by the interim cohort projections.
The principle purpose of the new CMI model appears to be as a kind of "common currency" for offices to express their projection bases. In this sense it is intended as a replacement for the commonly used interim cohort projections, which are indeed due for replacement as they are based on data up to 1999. However, the risk is that some people will see this new model as the CMI's recommended projection tool. A further problem is that the tool can generate far too many alternative scenarios for it to function as some kind of standard or "common currency". Perhaps to that end it would have been more effective to simply update the three interim cohort projections?