## Seasoned analysis

The importance of seasonal analysis was underscored by a recent letter form the UK insurance regulator.  In a previous blog, I looked at quarterly seasonal variation in a portfolio of defined-benefit pensions, and in a more recent blog I looked at monthly seasonal variation in mortality in England & Wales. However, rather than split observation periods into ever-smaller units, why not analyse seasonal mortality continuously? And if we can model it continuously, can we detect seasonal variation in any portfolio?

At this point you could be forgiven for thinking that modelling seasonal mortality were irrelevant for underwriting pension-scheme buy-outs or longevity swaps. After all, any fluctuating effects will even out over the long-term for a portfolio of pensions in payment. However, there are some subtle potential benefits of allowing for seasonal variation:

• Enhancement, i.e. the phenomenon whereby the inclusion of one statistically significant risk factor could improve the explanatory power of the other factors in the model.
• Flexibility.  Allowing for seasonal effects liberates the analyst from ensuring that an exposure period has equal numbers of each season. This is particularly useful when analysing portfolio experience for a bulk annuity or longevity swap, where the supplied data often cover a non-integral number of years and it is important to use all available data.

We can allow for seasonal effects in continuous time by adding a simple cyclic term to any existing model for mortality, $$\mu_{x,r,y}$$, as follows:

$\log\mu^*_{x,r,y} = \log\mu_{x,r,y} + e^\zeta\cos\left(2\pi(y-\tau)\right)\qquad (1)$

where $$x$$ denotes age, $$r$$ denotes duration and $$y$$ denotes calendar time. $$\tau$$ represents the proportion of the year after 1st January when mortality peaks and where $$e^\zeta$$ is the peak additional mortality at that point. There is an interactive online tool here to demonstrate this.

Since equation (1) is an addition to $$\log\mu_{x,r,y}$$, the mortality hazard will be multiplied by a seasonal factor that fluctuates smoothly and continuously around 1. Table 1 shows the estimated seasonal mortality parameters for different portfolios in Northern Europe, where mortality tends to peak between late December and early February (seasonal effects are shifted by six months in the Southern Hemisphere):

Table 1. Seasonal peak mortality for selected international portfolios. Source: Richards (2019) plus own calculations.

CountryPortfolio nature$$\hat\zeta$$$$\hat\tau$$Peak as
% of
average
Peak time
of year
Scotland Pension scheme -1.88 0.082 117% Jan 30th
UK Insurer annuities -2.00 0.001 114% Jan 1st
Canada Pension plan -2.04 0.109 114% Feb 9th
England Pension scheme -2.02 0.071 114% Jan 26th
Netherlands Pension scheme -2.25 0.052 111% Jan 20th
France Insurer annuities -2.42 0.066 109% Jan 25th
USA Pension plan -2.63 0.074 107% Jan 27th

Table 1 shows that seasonal effects can be consistently detected in almost any portfolio — the Scottish pension scheme in the first row has fewer than 18,000 lives, while the UK annuity portfolio in the second row has just four years of experience. Seasonal effects are strong, hence they are relatively easy to detect using equation (1).  Indeed, we recently went further in Richards et al (2020) and showed how seasonal variation increases with age, and is larger for low-income pensioners.

References

Richards, S. J. (2019) A Hermite-spline model of post-retirement mortality, Scandinavian Actuarial Journal, DOI: 10.1080/03461238.2019.1642239.

Richards, S. J., Ramonat, S. J., Vesper, G. and Kleinow, T. (2020) Modelling seasonal mortality with individual data, Scandinavian Actuarial Journal, DOI: 10.1080/03461238.2020.1777194.

### RECENT POSTS

In my previous blog I showed how suddenly the excess ... Read more
According to British Prime Minister Harold Wilson , " a ... Read more
In building a mortality model (or any other kind of ... Read more
Stephen Richards is the Managing Director of Longevitas
##### Seasonal patterns in Longevitas

Longevitas supports two methods of modelling seasonal patterns:

1. The CalendarPeriod variable, and
2. The SeasonalEffect variable.

Longevitas users can fit models with a variety of period effects using the CalendarPeriod variable. Simply go to the Configuration section and enable this in the Modelling tab. There you will also have the option to select the frequency of effects, as well as their alignment during the year. The CalendarPeriod is a categorical variable, i.e. the effect is assumed to be constant within the period.

The SeasonalEffect variable is a feature of the Hermite family of models. It is a continuous variable, and so is a more parsimonious option when modelling post-retirement mortality differentials.