A DYNAMIC
MODEL FOR PENSION FUNDS MANAGEMENT
361
A DYNAMIC
MODEL FOR MANAGEMENT
PENSION
FUNDS
JEAN-FRAKCOIS BOULIER DIRECTHJRD~LAFCECHERCHEETDEL'INNOVA?‘ION ETIENNE TRUSSANT INGENIEURFINANCIBR CCF,103 AVEN~D~SCIIAMI'SELYS~~S 75008 PARIS DANIELE FLORENS PROFL~SEUR UNIVERW~DEPARIS IX DAUPMINE TEL. : 33 140 70 32 76 FAX : 33 140 70 30 31
ABSTRACT The management of pension funds financial encompasses asset allocation and the control of the future flows of contributions. A high proportion of stocks in the portfolio has the benefit of a lower mean contribution level, but at the price of a higher time variation of contribution flows. This paper models the trade-off in a inter-temporal framework and uses of stochastic control to obtain an optimal asset allocation -between a risky asset and a riskless asset- and the contribution policy. The solution in the case of a defined benefit scheme shows that the proportion of the risky asset and the level of contribution are both proportional to the difference between the maximum wealth necessary to fund all pensions, and the actual wealth of the pension fund. Illustrative simulations for France, US and Japan for various periods show a decrease of the contribution level down to zero after some decades. The hypothesis made and some shortcomings of the model are discussed and further research is outlined.
A DYNAMIC
MODEL FOR PENSION FUNDS MANAGEMENT
363
I Tntrocluction The economic role of pension funds is considerable and well acknowledged even in countries where the pay-as-you-go systemis dominant. Indeed, at their mature stage, which they are at in many Anglo-Saxon countries, pension fund assetsrepresent a large percentage of stock and bond market capitalisation, in size comparable with the country’s GNP, and the contribution flows to the pension fund account for a significant part of personal savings. Moreover, their final objective -to pay pensions to workers in their old age- is of upmost importance for the social and political stability of the wealthy economies. How to manage pension provisions adequately is thus a crucial question. The principles underlying pension funds are quite simple, even if the variety of actual schemesfrom one country or one industry to another is vast and complex. Workers and corporations pay contributions to a pension fund, which invests them over a very long period of time and releases them when the workers retire, in the form of pensions. Obviously, the more the contribution, the higher should be the pension. Nevertheless, asset allocation also comes into play, in so far as that even a slight improvement of the asset portfolio mean return, say one or two percent, may result after thirty years of accumulation, in a sizeable increase -by 40% to lOO%- of the pensions. On the other hand, too much exposure to stock market fluctuations could, in the absence of careful management of the asset portfolio, severely damage the assetvalue and impose an undesirable increase of the contributions. In this conlcxt, portfolio management and the contributions scheme are clearly interdepcndant. Moreover, the decisions made over one year have certainly consequences the future. Therefore multiple horizon optimization seemsto in be appropriate. Because stock returns are uncertain in efficient markets, stochasticcontrol would help in finding the optimal investment policy, as well as the adequatelevel of contribution. Up to our knowledge, the use of stochastic control for pension fund financial management has not been reported in the literature. However, Met-ton (1972) described the basic framework of intertemporal optimization and showed how the Bellman function can provide a solution to the asset allocation of an investor, given his objectives and risk tolerance. On the other hand, asset and
364
5TH AFIR INTERNATIONAL
COLLOQUIUM
liability management has been invoked by scvcral rcccnt sludics in order to determine the asset allocation of pension funds. Sharpc and Tint (1990) proposed to optimize A-kL, where A and L stand for asset and liability rcspectivcly and k is a positive constant less than enc. Surprisingly, as remarked by Sharpe and Tint, investment policies of pension funds was hardly a subject of interest bcforc the early 90’s. A few authors, such as Tcppcr and Aflleck (1974) and Black and Jones (1988) had mndc attempts to propose solutions to the asset allocation problem, given the liabilities of the pension fund. Nevertheless as asset and liability management tcchniqucs improve and arc put into practice, an increasing number of papers have addressed the issue, now seen as important. Among them, Leibowitz ct al. (1993) showed how to cope with a number of conflicting constraints and to come up with an appropriate and yet simple optimization. Griffin (1993) has also presented a methodology which he has applied to Dutch and English pension funds globally, in an attempt to explain why the Dutch invest less than 25% in stocks, whereas the British invest more than 80%. All thcsc studies stick to the asset allocation problem in a one period framework. However Bocnder et al. (1993) have tried to investigate the linancial management problem in a more general setting, making USC of a scenario approach. The outline of this paper is the following. The next section describes the financial framework and poses the optimization problem. The solution of the problem is presented in the following section, in the case of dcfincd bcncfits. The last section discusses the results from a financial and economic point of view, and examines some possible generalizations and improvements of the method.
A DYNAMIC
MODEL
FOR PENSION
FUNDS
MANAGEMENT
365
II Financial setting 1. Definition of variables In rhc rest of the paper the following variables will bc used :
Y,
ct Xf St
4
pension payments contributions portfolio market value market value of the risky asset investment in the risky asset,as a proportion of portfolio value.
All of them arc functions of time t, either stochastic or deterministic. On the other hand, the following variables are supposed to be constant : risk free rate risk premium (positive) volatility of the risky asset pension growth rate psychological discount rate contribution growth rate
cs a B Y
Finally, WC shall refer to (c*, u*) as the optimal policy and to xm as the maximum ncccssarywealth.
2. Main hypothesis. We consider the financial management of the aggregated pension fund position and assume that either pension flows or contribution flows in the future are known. The first case corresponds to the defined benefit type of pension fund and the second to the defined contribution type. For sake of simplicity, their growth rates are taken as constant, a and y respectively. This growth rate may account for a demographic trend, an inflation scenario, a purchasing power evolution or any kind of combination of these factors as
366
5TH AFIR INTERNATIONAL
COLLOQUIUM
long as it lcads to a deterministic growth ram. This last hypothesis is certainly an important limitation of the model which will bc discus& further. Thus, in the dclincd contribution pension fund WChave :
and for the dclincd contribution
pension fund :
dc, = l/c,
In each cast, the contributions made arc invcstcd in a portfolio allocated in various financial assets. Although more general assumptions arc clearly possible, we restrict this study to the case of two assets invcstigatcd by Merton (1972) : -a riskless asset whose return is r, the risk-free rate, assumed constant ; -a risky asset whose price S, follows the standard geometric brownian motion
dS, IS, =(h+r)dt+odW,
where Ltw, denotes the usual differential of brownian motion. The expcctcd return of this risky asset, r + h, is therefore higher than the riskfrce rate, the difference being the constant risk premium ?L. On the other hand the future returns of the risky asset are not known with certainty because of the volatility (assumed to bc constant) and the stochastic process W. In this simple setting the portfolio managcmcnt consists in allocating a proportion 1~~of the value xt into the risky asset. Typically the portfolio is composed of stocks and bills. Again, gcncralization is possible and will bc discussed later.
3. Optimization
As mentioned before there are basically two cases. In the defined benefit framework one should try to minimize the contributions, with the obligation to meet the liabilities of the pension fund. On the other hand, the defncd
A DYNAMIC
MODEL
FOR PENSION
FUNDS
MANAGEMENT
367
contribution pension fund manager aims at maximizing the pensions to bc paid to the retirees, knowing the stream of contributions. In this study, we concentrate on the defined benefits case, which occurs, more frequently leaving the development of the defied contributions caseto another study. WC suppose that contributors are reluctant to pay higher contributions either today or in the future, but that they have their own judgement as to the discount rate which we have denominated the psychological discount rate p. For sake of mathematical tractability we have also assumed that their disutility is a power function of the contribution c. In the rest of the paper the exponent will be taken as 2, but generalization is possible. Under these circumstances,a rational pension fund manager would try to minimize :
V = exp(-ps)c,2 ds
I 0
The optimum policy must satisfy the following constraints : -payment of the pension y, -positive value for xt Therefore we seek the policy (cl, LQ which minimizes E(V) under the two preceding constraints.
III
Solution
ct and I+ being respectively the contributions and the pensions paid per unit of time continuously discounted, the evolution of the portfolio is described by the following equation :
368
5TH AFIR INTERNATIONAL
COLLOQUIUM
The first term on the right is the growth of wealth due to the part of the portfolio invcstcd in the risky asset. The second term eomcs from the part invested in the risklcss asset. The third term rcprcscnts the flow due to the balance of subscriptions and payment of pensions. Making USCof the model assumed for S, the above equation can be rcwrittcn as :
dX, = [ rX, + hu, X, + cl - pt ]dt + u, X,odlfl(
For technical reasons, we assume that
O=inf.(e-P’c?+V;+V~m+V’~d~~+~~,~”,,(d~)”,
Using the equations above we have :
2r--p-i?/G2 >O
In the cast where p = r, this assumption becomes
r--h2 I 02>0
This high If V( then
inequality means that the risk premium is to bc justilicd by a suflicicnt volatility (CT> X/L). t ,x,p) denotes the value function of the problem, Bellman’s equation is :
prc2 +V’t+(rx+3LU.x+C-py)V’x+cxpV’,,+~V”,,, .\:*u*o 21 (1)
0 =inf(C
under the constraints x > 0 et u > 0. The term in brackets is a polynomial function of U and C , therefore the optimal policy (u*,c*) satislies
A&xv +V’& X2ULcJ2 0 ; =
That is to say
A DYNAMIC
MODEL FOR PENSION hV;
FUNDS MANAGEMENT
369
U” =-
v ‘lx x XI2 * c =-v;
(2)
2e-bt
(3)
Substituting the expressions (2) and (3) in (1) leads to
1 t2 -‘;?e pt V.+V;+(r~-p)\/;+w~V’~---J2 v;” 202 V”,., =o
Let us now search a priori for a solution of the type
V(t,x,p)=e?F(x,p)
The differential equation satisfied by F is
-$F’z-PF+(rx-p)F’x+apF’,--h2 Ff2 20~ F”,,, =o
Remark that this equation is homogenous for the variable y = x/p. then set F(x, PI = P2f(dP)
Let us
The differential equation satisfied by f is now
Once more, the solution is obtained by searching a priori for a solution of the form :
370
5TH AFIR INTERNATIONAL
COLLOQUIUM
J.(y)=Ay2+By+C We identify A, B and C and tind
f w=iZr-P-p-&$
x2
Combining the equation for V ,F and the last one, WC find the final expression for the value function :
V(t,x,y)=e-Pf(2r-p-A2/~2)(X--/(I.-a))2 (4)
Direct substitution in (1) shows that this function is cffcctively the solution of the problem in the domain 0 I x 5 x,, where x,~ = II/( r - (Y). In the domain x, I x, it is obvious that the optimal policy is zero-contribution and no risky asset in the portfolio. Using (2), (3) and (4), the final expressions for the oplimal policy arc :
c* = 1 (2r-p-P/d)(X, 0 -x) if s<x, if x, I s
The condition
CI > h&
ensures that the constraints on u and C are satisfied.
A DYNAMIC MODEL FOR PENSION FUNDS MANAGEMENT
IV Discussion
371
l- Some comments
from an economic
point of view
The value x, represents a critical threshold equal to the discounted value of all future flows due to pension payment, for an infinite horizon. Above it, the fund is entirely invested in riskless assets and receives no contribution. It is the situation of a rent : the interest generated by the principal is suflicient to pay the pensions. Below this threshold, the fund is partially invested in risky assets and receives contributions. The optimal contribution (c*) and the amount invested in risky assets (u*x) are decreasing linear functions of the fund’s value when it is under the critical threshold. The general idea is that the wealthier the fund, the less risk it needs to take, and the more cautious and risk-adverse it appears.As a consequence,the pension fund tends to buy more stocks when the market falls. Moreover as we have not contrained u , it can be greater than one, meaning that the fund will borrow in order to invest more in stocks. Although one can expect such a behaviour for a very long term investor it is certainly necessaryto set some limits to prevent too large a debt. In addition, a minimum investment in bonds could be imposed in order to ensure the payment of the accumulated liabilities to those workers who have contributed. Note that, logically, the higher the volatility and the lower the risk premium, the higher the contributions and the lower the part invested in the risky asset. In this simple framework there is no defined time horizon and thus the fund manager knows he can benefit from the higher return of stocks. His incentive to do so increases obviously with the risk premium. As it is also a decreasing function of o2 one could consider I / o2 as a characteristic time for the investor. The longer this time horizon the higher the propensity to invest in stocks.
2. Historical
simulations
For sake of illustration a number of simulations have been performed, making the assumption that recent historical data lit out simple model. Making use of
372
5TH AFIR INTERNATIONAL
COLLOQUIUM
early records of inflation, stock market prices (in form of a domestic market index with reinvested dividends), and inlcrcst talcs (short tcnn) for France, Japan and the USA, we have computed the pararnctcr values shown in table 1 below.
h France Japan USA
1960-92 1973-93 1970-93 9 7 8.9 2.4 2.9 4.3 19.5 17.5 15.7
0.
6.4 5.1 5.9
r-(hi
6.5 4.3 1.5
o)2
Table I Parameter values (%) used in the simulation Note that these data correspond to different time periods (see Column 2). Indeed the performance of the strategies will largely dcpcnd on the stock market returns over the first years of simulation. Figure 1 shows the stock index evolution for the markets considered. Two pension funds were considered . PFl has a comfortable surplus of 50% of its curncnt accumulated liabilities, whcrcas PF2 has a limited smplus of 10%. This surplus is by definition the difference between the actual portfolio value minus the discounted liabilities at the rate r. For sake of simplicity WC have assumed that these liabilities were equal to all the pensions to be paid over the next thirty years. We have simulated yearly contributions and reallocations of the portfolio according to the optimal policy derived in section III. WC have not studied the sensitivity of the result to the time interval between contributions and reallocation, but it would certainly have some influence. The resulting evolution of the portfolio values, of the proportion (%) of stocks held in the portfolio and finally of the contributions arc rcprcscntcd in Figures 2 a, b, c (respectively 3 and 4) in the cast of France (respectively in the case of Japan and USA). The long term benefits of the solution are visible in the French case. No more contributions are necessary after 23 years for PFl and 32 years for PF2. The
A DYNAMIC
MODEL FOR PENSION FUNDS MANAGEMENT
373
decreasing exposure to the stock market is striking. Howcvcr for both there is borrowing in the first year. This situation is also encountered in the case of the Japanesepension funds and to a much larger extent in the American case. Nevertheless in this latter case the upper bound (x,,, 1 is reached within a short time period of 11 years for PFl, although it entails a very high leverage. On the other hand, none of the Japanese pension funds reach the point at which no more contributions are necessary.The impact of the Tokyo markets poor performance after 1990 is clear. The pension fund is obliged to raise the contribution and increase at the same time its exposure to the stock market.
3. Criticim and further research This too simple model suffers from a number of shortcomings that we have already mentioned. Let us try to summarize them and to discuss the possible developments that could overcome these difficulties. First, the investment in stock u was not bounded (to 1 or to some level depending on the liabilities). Adding such a constraint is obviously not a theoretical problem, becauseunder this situation the solution would still exist. But it will no longer allow us to derive a closed form solution. Numerical methods must be used, the accuracy of which can fortunately be checked in the non-constrained case. Secondly the condition 2r - p - (3L/ o I2 > 0 need not be met, as when the stock market does not achieve a reasonable return-to-risk ratio. Such a situation could force the pension fund to look for a better expected performance in foreign markets. If p remains still too big compared to the available foreign market performance, probably a pension funds system cannot be achieved, simply because the aversion to savings is too high. Thirdly, the single risky asset could without theoretical difliculty bc replaced by a much more realistic mix. A finite-time horizon corresponding to known mortality tables (in average of course) could also be included. In this case the derivation of the close form solution will bc cumbersome, though possible.
374
STH AFIR INTERNATIONAL
COLLOQUIUM
Inllation was considcrcd to bc of constant rate, which is an unrealistic assumption, cspccially in Europ.xm countries, whcrc inflation over a long period of time happens to have been devastating. A stochastic description of inflation would have to consider stochastic real rctums of various asset classes, with the possible problem of having serial correlations to take into
account.
Finally, the economic criteria chooscn for the optimization should bc rcconsidcrcd. Indeed the variation in the contributions is probably unbcarablc for economic agents who would prcfcr a smoother pattcm cvcn at the cxpcnsc of a higher mean contribution rate. In addition, no risk aversion was considered in the portfolio, which is also unrealistic. Again, general criteria would lcad either to no solution or to a numerical solution with some real computing difliculties. Results pertaining to the defied contribution cast have already been obtained and will be pcscnted in another paper. The framework remains simplistic in the interest of obtaining a closed form solution.
A DYNAMIC MODEL FOR PENSIONFUNDS MANAGEMENT REFERENCES
375
Fisher Black and Robert Jones, 1988, Simplifying portfolio insurance for corporate pension plans, J. of Portfolio Management,Summcr. C.G.E. Bocnder, C.L. Dert, P.C. van Aalst, D. Barns and Hecmskerk, 1993, Scenario approaches for asset-liability management, Inquire, April. Mark Griffin, 1993, A new rationale for the different asset allocation Dutch and UK pension funds, 3rd AFIR Conference, Rome, April.
of
Martin L. Leibowitz, Stanley Kogelman, and Lawrence N. Bader, 1993, Asset performance and surplus control : A dual-shortfall approach, J. of Portfolio Management,Winter . R. Merton, 197 1, Optimum consumption and portfolio rules in a continuoustime model, Journal of Economic Theory, 3, (p 373-413). William F. Sharpe and Lawrence G. Tint, 1990, Liabilities approach, J. of Portfolio Management, p. 5-10, Winter.
- A new
Irwin Tepper and A.R.P. Affleck, 1974, Pension plan liabilities corporate financial strategies, The Journal of Finance, December.
and
376 APPENDIX
5TH AFIR INTERNATIONAL
COLLOQUIUM
Figure 1
lzll -ml -833 co)-a-m--
I
I
A DYNAMIC
MODEL
FOR PENSION
FUNDS
MANAGEMENT
377
Illustration for a French fund
I:igure 2a : The value of the fund
4500 4000 3500 3000 2500 2000 1500 1000 500 0 1960 1965 1970 1975 1980 1985 1990
Years
-m-
fund no1
---IS--
fund no2
Figure 2b : Percentage
140% 120% 100% 80% 60% 40% 20% 0% 1960 1965 1970
invested in risky assets
1975
1980
1985
1990
Years -Hfund no1 ---Ck-fund no2
378
Figure
5TH AFIR INTERNATIONAL
2c : Contributions
COLLOQUIUM
SD-¤-¤DDDD 1975 1980 1985 1990
Years
ADYNAMICMODELFORPENSIONFUNDSMANAGEMENT Illustration for a Japanese fund
379
Figure 3a : The value of the fund
1600 1400 1200 1000 800 600 400 200
T
1973
1975
1977
1979
1981
1983
1985
1987
1989
1991
1993
Wealth
Years
-Ifund no1 + fund no2
380
5TH AFIR INTERNATIONAL Figure 3b : Percentage invested in risky assets
COLLOQUIUM
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
Proportion
of risky asset
Yeas
--Ifund no1 -Dfund no2
A DYNAMIC
MODEL
FOR PENSION
Figure
FUNDS
MANAGEMENT
381
3c : Contributions
16 ‘---’ :; I ,/” -1W--1---,-,-, ,A’\’ \ ‘--I \ I 10 -. a -6 -. 8
1974
1976
1978
1980
1982
i 984
1986
1988
1990
1992
Amount
of contribution
-I-fund
Years
no1 + fund n”2
382
STH AFIR INTERNATIONAL
COLLOQUIUM
Illustration
for a US fund
Figure
4a : The
value
of the funds
1970
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
Wealth
Years
A DYNAMIC
MODEL FOR PENSION
FUNDS MANAGEMENT
383
Figure 4b : Percentage invested in risky assets
384
5TH AFIR INTERNATIONAL
COLLOQUIUM
Figure
4c : Contributions
Amont
of contribution
1971
1973
1975
1977
1979
1981
1983
1985
1987
1989
1991
1993
Years
-a--fund no1 --C--fund n”2