Case study: policy function iteration for a cohort of clients

5.2 Decumulation phase

5.2.3 A stationary solution

5.2.3.4 Case study: policy function iteration for a cohort of clients

We finally demonstrate the presented policy function iteration algorithm in a numerical case study, where we concentrate on the cohort perspective. First, the optimal investment decision variables are determined for every state under the infinite-horizon problem (stationary solution). Afterwards, a simulation analysis in the finite-horizon model, where the approximate optimal stationary solution to the infinite-horizon model is applied, provides the most relevant numbers and probabilities and compares the considered strategies.

For ease of comparison, we select the very same grids and parameters as in the former case study in Section 5.2.2.5. Furthermore, we determine and consider the optimal investment strategy for α “ p0%|20%|40%q (no | moderate | pronounced buffer).

Optimization. Let V0 “10, 000 as before and let us define the state space grid by Vmin“20%ˆV0, V_max “ 500% ˆ V0, nV “ 1, 000 for GridpV q. We further select a step size of 1% for GridpCCRq, thus nCCR“26. This leads to a total grid size of nS “26, 000 states. The grids for Z and a remain unchanged. In particular, Assumptions 5.12 and 5.14 are fulfilled. We seek for a fixed-point solution to the value function according to the policy function iteration algorithm in Section 5.2.3.3. We would like to comment that it only takes seven iterations maximal (niterď7) to find the fixed point for α “ p0%|20%|40%q and with that the stationary solution to the infinite-horizon optimization problem. Thus, the algorithm converges very quickly (takes « 2h for each α).

Figure 5.17 visualizes the average optimal risky relative asset allocations ˆπ^‹(inv) “ a^‹ (investment portfolio) and ˆπ^‹(total) (total cohort portfolio) for all CCR^(total)c values in the grid. The grid of the state space was constructed such that there are nV “ 1, 000 different V values and nCCR “ 26

26Notice that λx“ 1.18% is suitable if one looks at very short planning horizons. Since in our case the planning horizon is artificially infinity, we can alternatively calibrate the mortality rate λx towards the remaining life expectancy of a 65-year old client such that it coincides with the average remaining expected lifetime of 21.06 years (female) and 17.87 years (male) in Germany, cf. Statistisches Bundesamt (2019). We receive λx“ 5.14%. Then the convergence factor drops to e^´pλ^x^`βq∆ “ 0.9218 after one iteration (step size ∆ “ 1 year), e^´pλ^x^`βq10∆ “ 0.4431 after ten iterations and e^´pλ^x^`βq100∆“ 0.0003 after hundred iterations. This shows that a higher mortality rate leads to a faster convergence of the algorithm.

27In practice, instead of calculating the inverse matrix one typically rather solves the system of linear equations associated with Eq. (5.68) to obtain V^piqpSq in a more efficient fashion.

(a) a^‹. (b) ˆπ^‹(total).

Figure 5.17: Average a^‹ and ˆπ^‹(total) for a given CCR^(total)c value in the grid.

different P values for every V value selected such that the corresponding grid for the CCR has a step size of one percentage point. Hence, for every CCR P t100%, 101%, . . . , 124%, 125%u, we build the average over the nV values for a^‹ and ˆπ^‹(total) that have an equal CCR value. The pattern comes close to an S-shaped form: It can be seen that a higher buffer parameter α, in particular for α “ 40%, leads to a lower relative risky investment for small CCR^(total)c values, but catches up for large CCR^(total)c values. This is a desired behavior, since it implies a lower risk of a pension shortening for small CCR^(total)c values within the range r100%, 110%s, without losing the upside potential of a pension enhancement for CCR^(total)c values close to 125%. Furthermore, except for the region CCRc^(total)P r100%, 105%s, the average optimal risky relative investment increases with the CCRc^(total) value. This is meaningful since with a higher CCR^(total)c value, one is less exposed to the risk of falling outside the left boundary of the CCR^(total)c corridor (pension reduction risk).

The higher risky investment close to 100% is also reasonable. Imagine the CCRc^(total) is close to 100%; if now the risky allocation is very small, even some positive return of the underlying asset class cannot compensate for the outflows (cohort-related pensions), which pushes the CCR^(total)c

below 100% with a high probability.

Simulation study. We next carry out a brief simulation study with a longer time horizon of T “ T `˜ 10∆ “ T ` 10 years (∆ “ 1), i.e. simulation over 10 years compared to the two years in Section 5.2.2.5. We start with the same initial states S0 “ pV0, P₀q “ p10, 000, p277|270|258qq for α “ p0%|20%|40%q; thus the initial CCRc^(total)pT q and the initial buffer-to-wealth ratio ^V^c^(buffer)_V₀ ^{pT q} as well as the investment-to-wealth ratio ^V^c^(inv)_V₀^{pT q} keep the same. We simulate 10, 000 paths of the relevant processes where we use the optimal stationary solution as asset allocation that corresponds to the closest grid point.

We assume that the average mortality (explained in Section 5.2.2.3) for the cohort is realized.

We look at the optimal relative pension evolution ^P^‹^ptq

e´λxpjqpt´T qP0, where P^‹ptq denotes the cohort pension at time t under the optimal stationary asset allocation strategy a^‹ “ a^‹pSq. Notice that

P^‹ptq e´λxpjqpt´T qP0

“ ^P^‹^pt`∆q

e´λxpjqpt`∆´T qP0

indicates a stable individual pension for the customers in the cohort from time t to t ` ∆, i.e. if ^P^‹^ptq

e´λxpjqpt´T qP0 is stable then the individual pensions are stable. In what follows we always look at the individual pension perspective in the cohort. Moreover, let V^‹ptq denote the total cohort wealth at time t under a^‹ “ a^‹pSq, in what follows we look at the optimal relative total wealth evolution ^V_V^‹^ptq₀ .

We would like to point out that the pension P^‹ptqand the asset allocation decisions a^‹ptq, ˆπ^‹(total)ptq are constant on every annual interval rt^piq, t^piq`∆q and are only changed at the evaluation times t^piq.

Table 5.7 illustrates the probabilities of pension shortenings and enhancements. Table 5.8 provides risk and reward numbers for the relative pension and the total wealth. In general, we observe that a higher buffer parameter α significantly improves the probabilities in Table 5.7 from a client’s perspective. In particular, the probability that the average individual pension that is to be paid out over the entire period is larger than the initial pension level P0 and the probability that there are more pension enhancements than reductions are quite high, especially for α “ 40%. However, both the (relative) risk in terms of volatility and Value-at-Risk and the (relative) reward in terms of expected value do not suffer, which is remarkable. Actually the opposite is the case: A higher buffer parameter α leads to a higher average of the relative pension level and a lower standard deviation (lower standard deviation of relative pension means a more stable pension development);

moreover, the worst case relative pensions in the tail (Value-at-Risk) also exceed the ones for smaller α. The single exception is the volatility of the pension, where α “ 20% shows a slightly smaller number than α “ 40%. Those benefits of the α ą 0% portfolios comes at the cost of an initially lower pension level P0 “ P0pαq, which represents a tradeoff between the initial pension level and future pension properties. The selection of the case-specific optimal α value, named α^‹, depends on the respective target or criterion. If for instance the probability of at least one pension shortening shall coincide with a pre-defined probability pred, α^‹ can be selected such that the corresponding probability comes closest to pred. Alternatively, α^‹ could be selected such that the expectation of the sum of pension cash flows gets maximized.

In summary in terms of the individual cohort pension (relative to the initial pension level P0), one can see that α “ 40% outperforms the α “ p0%|20%q strategies, and the α “ 20% outperforms the α “ 0% strategy. The higher the buffer parameter α, the more the downside risk is limited, and even the upside potential is enhanced.

We draw the conclusion that our proposed model, where we divide our total wealth into an invest-ment and a buffer portfolio, leads to a sophisticated optimal dynamic asset allocation policy that is performance seeking while reducing downside risks and improving probabilities; hence provides remarkable and meaningful benefits to clients.

Finally, we simulate the optimal strategy a^‹, the pension P^‹ and the wealth V^‹ evolution under three different scenarios: a bullish, a bearish and a non-directional market. In each simulation we need to generate the risk driver Z for every period. Figure 5.18 provides the corresponding underlying risky asset class price processes, denoted by VZptq, that correspond to the development of Z. Next, Figure 5.19 illustrates the evolution of the relative pension, Figure 5.20 visualizes the very same but for the total wealth. From Figure 5.19 we infer that

1. the individual pensions increase more often for higher α and even end up with a higher terminal pension (relative to P0) in a bullish market,

Probability²⁸

Pp“at least one pension reduction”q: 49.5%|36.4%|25.5%

Pp“path-wise average pension ě P0”q: 74.1%|80.8%|86.1%

Pp“number of pension enhancements ě number of pension reductions”q: 84.4%|91.3%|96.8%

Table 5.7: Probabilities of pension rate changes for α “ p0%|20%|40%q.

29 α “0% α “20% α “40%

” P^‹ptq e´λxpjqpt´T qP0

ı: 107.5% 108.4% 110.9%

´ P^‹ptq e´λxpjqpt´T qP0

¯: 18.1% 16.8% 17.0%

V aR_0.05

´ P^‹ptq e´λxpjqpt´T qP0

¯: 83.3% 84.8% 86.5%

V aR_0.01

´ P^‹ptq e´λxpjqpt´T qP0

¯: 71.9% 75.1% 77.0%

”V^‹ptq V0

ı: 96.6% 96.2% 96.5%

´V^‹ptq V0

¯: 15.7% 14.3% 14.6%

Table 5.8: Relative performance numbers for α “ p0%|20%|40%q under 10, 000 simulations.

2. the individual pensions decrease only once for α “ 40% but twice for the remaining (α “ p0%|20%q) in a bearish market,

3. and the individual pensions do not decline for α “ 40% but do decrease and behave very unstable and volatile for the remaining (α “ p0%|20%q) in a non-directional market.

In total, the number of pension reductions for α ą 0% (with buffer) never exceeds the respective number for α “ 0% (no buffer) in the considered representative scenarios.

Figure 5.21 complements the former figures on the pension and wealth evolution with a visualization of the CCR^(total)c ptq development. While the CCR^(total)c ptqvalues for α ą 0% (with buffer) do not generally fall short the respective values for α “ 0% (no buffer), the α ą 0% portfolios need less pension shortenings to keep the CCR^(total)c ptqinside its target corridor. Therefore, with selecting a higher α% value, one can improve the management of the wealth such that the CCR^(total)c ptq remains more stable in its corridor without reducing the pension.

In addition, Figures 5.22 and 5.23 show the optimal asset allocation policies a^‹ptqfor the investment wealth and ˆπ^‹(total)ptqfor the total wealth. One can observe that the optimal strategy for α “ 40%

frequently behaves opposed to the optimal strategy for α “ 0%. Moreover, Figure 5.24 illustrates

28We count the number of paths which fulfill the statement in Pp¨q. The relative frequency then serves as an estimation for the probability.

29Note that ^P^‹^ptq

e´λxpjqpt´T qP₀ “ 1 and ^V_V^‹^ptq

0 “ 1 for initial time t “ T .

the kernel density estimates for the path-wise average pensions and wealths. Note that for one path, a higher path-wise average pension automatically implies a higher total sum of pension cash flows received by the customer. The figure points out that although the distributions of the wealths are rather close among all considered α values (see also expected values and volatilities in Table 5.8), the distributions of the relative pensions differ. The pension distribution for α “ 40% has lower probability on the left end and is more shifted to the right; this is also reflected in Table 5.8. Thus, a pension fund client that follows the α “ 40% strategy benefits in terms of the pension distribution since lower pensions compared to the initial pension level P0 are on average less likely.

However, as already explained, these benefits come at the cost of an initially lower pension level P0. We would like to comment that the averages over all simulated a^‹ptqand ˆπ^‹(total)ptqvalues are very close to each other among the three considered buffer parameters α. However, as analyzed above, the relative performance and characteristics of the optimal portfolios with a buffer (α ą 0%) are superior over the optimal portfolio without a buffer (α “ 0%). This shows that the dynamics and the structure of the asset allocation plays a crucial role.

Compared to the case study in Section 5.2.2.5 we now discover a generally lower level for the optimal risky relative allocations a^‹ptq and ˆπ^‹(total)ptq. However, the probability of one or more pension shortenings is now higher in comparison with Section 5.2.2.5. Both observations follow from the longer planning horizon: ten instead of two decision periods with a length of one year for every period. It is reasonable that the overall level of the risky relative allocation drops and the investment strategy is thus a bit more defensive if a longer investment horizon is considered because the investment return distribution becomes wider for a larger investment horizon and thus the probability that some barrier is crossed increases. If one aims for similar probabilities of pension shortenings, the risk thus needs to be reduced.