Mat- k 3.

2 Lecture Notes: Mathematical Modeling and Optimization with Applications in Finance

Soren S. Nielsen March 17, 1998

Contents

1 Introduction 2 The Dedication, or Cash-Flow Matching (CFM) Model.

2.1 The Deterministic Case . . . . . . . . . . . . 2.1.1 Short-term reinvesting and borrowing 2.1.2 Tradeability Considerations . . . . . . 2.1.3 Transactions costs . . . . . . . . . . . 2.1.4 Rebalancing an existing portfolio . . . 2.1.5 Time-mismatched cash ows . . . . . 2.1.6 End-of-horizon e ects . . . . . . . . . 2.1.7 Diversi cation considerations . . . . . 2.1.8 Net present values and duals . . . . . 2.2 A variation: Maximizing Horizon Return . . . 2.3 The Stochastic Case . . . . . . . . . . . . . . 2.4 Scenario Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

3 5 6 7 8 8 9 9 10 10 11 13

3 3

3 The Two-Stage Stochastic Program

3.1 Modeling future rebalancing: Dynamic models 3.2 A Formalization of the Stochastic Program . . 3.3 A Network Formulation . . . . . . . . . . . . . 3.3.1 Model De nition . . . . . . . . . . . . . 3.3.2 Extension to a Multistage Model . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

15

15 17 19 22 24

4 Multistage Stochastic Programs

4.1 Formulation of the T -stage Stochastic Program . . . . . . . . . . . 25 4.2 Scenarios and the Scenario Tree . . . . . . . . . . . . . . . . . . . . 25 4.3 The Split-Variable Formulation . . . . . . . . . . . . . . . . . . . . 27 5.1 5.2 5.3 5.4 5.5 5.6 Present Value and Dollar Duration. . . . . . The Immunization Model . . . . . . . . . . Macaulay and Modi ed Duration . . . . . . Critique of the simple Immunization Model Convexity . . . . . . . . . . . . . . . . . . . Factor Immunization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

24 28

5 Immunization: Present Value, Duration, Convexity

29 30 31 32 32 34

A Algebraic Formulation of Stochastic Network Model

36

1 Introduction
These notes support the Mat- k 3.2 course taught at the Department of Operations Research, University of Copenhagen, Fall 1997. The notes this year are just being written and are certainly in draft form. They are supplemented by 15]. The students in the class are being used as guinea-pigs and deserve thanks! For up-to-date information on the course, please visit the author's home page at
http://www.math.ku.dk/~nielsen/

and follow the course link.

2 The Dedication, or Cash-Flow Matching (CFM) Model.

Companies such as insurance companies often face a liability stream reaching several years into the future, for instance, representing future payouts on insurance products such as life insurance. Usually such future liability streams are stochastic, that is, it is not known today precisely when they will occur, or how big they will be. For a life insurance, this depends on customers' longevity and possibly on options built into the product, such as cancellation rights. Other examples of future liability streams are home owner mortgages (with xed or variable payments over a typically 15-30 year period), lottery payouts (for instance, the Texas Lottery pays out larger prizes in 20 annual installments), etc. It is often of interest to determine a portfolio of bonds (obligations) whose cash- ows replicate that of the liability stream. Regulatory committies often require insurance companies to demonstrate solvency, and one way to do this is to determine a \fair market value" of their liabilities by nding a replicating portfolio consisting of default-free bonds, such as Treasuries.This is similar to determining the present value (or \expected present value" in case of stochastic liabilities) of the liabilities using, for instance the zero-coupon yield curve, but is a more realistic measure because actually trading bonds, such as par bonds or even corporate or municipal bonds could also be used (although then credit risk needs to be addressed). Such a market value could also be used to sell the liability stream. This is done by many lotteries that pay out over many years but are not in the business of (or are prohibited from) money management.

2.1 The Deterministic Case

We consider rst the deterministic case, where assets as well as liabilities are known in advance. De ne: 3

T = f0 :::mg: The set of time periods, for instance measured in years, from t = 0 (\now") to t = m, the horizon. U = f1 ::: ng: The Universe of assets under consideration for inclusion in the
portfolio. Fi t : The cash ow arising from asset i at time t. Lt : The liability due in period t 1.

The cash ows Fi t can be both positive and negative by convention, a positive number indicates incoming cash and a negative number outgoing cash. For a bond which is purchased today at price Pi per unit face value, with an annual coupon payment of ci and a maturity of 4 years we would have Fi = (;Pi ci ci ci 1 + ci 0 ::: 0). The decision variables xi i 2 U denote the amount, in face value, of bond i that should be purchased. The objective of the model is to nd the least expensive portfolio whose cash ows at least satisfy the liabilities, and this cost is represented by the variable in this model: Minimize
x U
2 2<

CFM-1:
PF x + i0 i iP U F x
2

(Cash-Flow Matching Model) Subject to

i U xi
2

it

Lt t 1 0 I2U

The model works as follows, assuming positive liabilities and cash ows as shown above: Since portfolio cash ows in each time period must at least satisfy the liabilities the decision variables, or some of them, are forced to become positive. This in turn forces the cost, , to be positive (since the cash ows at time 0 are negative). By minimizing this cost, the least expensive portfolio is arrived at. It would seem natural to formulate this model using the objective Minimize

X
2

i U

Pi xi

remove the rst constraint above, and keep all cash ows positive elsewhere. This is a valid approach, but for instruments where cash ows may be either positive or negative (e.g., investments that are paid over several periods before returning to positive cash ows, or xed-for- oating swaps), there may not be a natural notion of an up-front price, so the formulation shown is somewhat more general. 4

Note that the liabilities themselves may be negative, and so may the optimal \cost". For this reason, in this model is often called a \lump" sum, to be paid or received up front. These possibilities come into play in the extended models given below. It is important to note that, even though several time periods are modeled, the CFM model is static in the sense that a single investment decision is made now, then not changed. In the context of a stochastic future (a redudancy...) one typically employs a truly dynamic model that allows for future rebalancing of the portfolio. In Section 2.3 we return to such Stochastic Programming models.

2.1.1 Short-term reinvesting and borrowing The only way for CFM to meet a liability in period t is for it to purchase instru-

ments with combined cash ows in that period to meet the liability. If the asset universe does not include an asset with a substantial cash in ow (for instance, a bond that matures) in period t, then instruments that mature later than t must be purchased in excessive amounts so their coupon payments can cover Lt . Even if a bond matures at t, it may be advantageous to purchase a much better priced bond that matures prior to t and simply deposit the money in the bank to meet the period t liability. Similarly, one could borrow, at the bank's rate, against a security which matures sometime in the future (in reality one might instead sell such a security, but the present model does not take such sales into account). Finally, short-term reinvesting and borrowing becomes natural with the integrality (tradeability) constraints we introduce later. Short-term reinvesting and borrowing can be modeled by adding variables rt and bt to represent the amount of cash reinvested from period t to t + 1, or borrowed in period t from period t + 1. There is usually no need to model multiperiod reinvesting or borrowing, since this is basically equivalent to a series of single-period acts. Assume that future, short-term investments have a period return of t , and that money can be borrowed at a rate of t . The augmented model is then: Minimize Subject to
x rb

(Cash-Flow Matching with reinvesting and borrowing)

CFM-2:

P F x +b + = r0 i0 i 0 iP U Fi t xi + (1 + t) rt 1 + bt = Lt + rt + (1 + t ) bt
2

i U bm xi ri bi
2

t 1 I 2U

=0 0

In this version cash can be carried between periods at the respective rates. Note that borrowing in the last period is, of course, prohibited. Reinvesting cash in the last period is not prohibited, but such cash is essentially wasted, although the variable rm can be interpreted as an end-of-horizon surplus. We have changed the inequalities in CFM's cash- ow balancing constraints to equalities since, with the possibility of reinvesting/borrowing there is no need for inequalities. The model does not usually, despite the increase in number of variables, become much harder to solve, but it is important to realize that we have introduced a risk that was not previously present in the model. This is in the assumption that we will in the future be able to borrow and reinvest at certain rates. If this assumption turns out not to hold, the model portfolio may actually generate a surplus or loss. In reality this is not usually a great concern as long as the reinvested and borrowed amounts are relatively small (which can be enforced easily using bounds), and the assumed rates are reasonable. However, it is customary to use quite conservative rates, such as t = 0 and t = 0:15 or higher. Note that reinvestment and borrowing rates need not be constant through time. They could, for instance, be a function of the planner's expectations of future interest rates.

2.1.2 Tradeability Considerations

Large institutional bond investors are usually interested in purchasing large blocks of individual bonds, in \round numbers" (\even lots") of face value. This is primarily because such blocks are more easily traded than smaller or \odd-lot" holdings, but also for liquidity reasons, to avoid the risk of getting stuck with a small holding of a bond with poor liquidity. In addition, brokers usually require a premium for \odd-lot" trades because they may have to split an \even lot" and incur a risk to sell the remaining odd lot. In short, we may need the model to allow only: 1. even-lot purchases, i.e., face-value amounts in multiples of, say, $100,000, and/or 2. minimum-lot purchases, i.e., purchase either none or at least some amount, say $500,000, of each bond. Even-lot purchases are naturally modeled using integer variables. If the decision variables xi denote for instance the number of $100 bonds to purchase and lots in multiples of $100,000 are desired, we just add integer variables yi and constraints xi = (100 000=100) yi 6

to CFM. Minimum-lot purchases, for instance $500,000 or nothing, are handled using binary variables, bi 2 f0 1g, and the constraints (500 000=100) bi xi and xi M bi where M is some large number such that, if bi = 0 then xi = 0, but if bi = 1 then xi (500 000=100) and essentially unconstrained upwards. When integrality conditions are used with CFM, it is important to also allow for reinvesting and possibly borrowing, even at very conservative rates, otherwise the model will be forced to overinvest and discard cash in most periods, which could result in a very sub-optimal solution. The above extensions both change the model from an LP to a Mixed-Integer Program, MIP, which can be extremely expensive to solve to optimality, even for moderate numbers of discrete variables. There are certain modeling tricks that may help (such as giving yi an upper bound as low as possible, and keeping M as small as possible, while still not in reality constraining xi ), but in practice it is often necessary to resort to heuristics such as rounding: Solve CFM in its LP version, then successively round variables to the nearest feasible value until a reasonable solution is found. Of course there is then no guarantee that an optimal solution has been found, but by comparing the value of a solution to the LP solution one can at least put a bound on the best possible improvement and thus estimate whether the solution found is \su ciently good".

2.1.3 Transactions costs

Associated with every investment is transactions costs, such as brokerage fees and bid-ask spreads. Certain model types (such as the Markowitz model, Section ??) are notorious for occasionally returning solutions containing small amounts of each of a large number of securities. Such a portfolio is in practice impossible to justify due to the costs of establishing and maintaining it. By explicitly modeling transaction costs, one can avoid this problem. We distinguish between two types of transactions costs: Fixed and variable. A xed cost is incurred for each instrument that is traded. This is modeled using a binary variable zi 2 f0 1g i 2 U for each instrument, and adding the term

X
2

i U

zi

to the objective, where F is the xed cost. These new variables need to be linked to the xi such that if xi > 0 then zi = 1. A set of constraints

xi M zi
7

where M is a suitably large number, takes care of this. The resulting program is, of course, a mixed-integer program, and can as such be very di cult to solve to optimality. Variable, or proportional, transactions costs, on the other hand, are easily modeled without any penalty in model complexity or di culty. A variable transaction cost is a per-unit cost, and this cost can simply be added to the cost of the instrument (or, more precisely, subtracted from Fi 0 in the CFM models). They arise, for instance, in modeling bid-ask spreads, in addition to brokers' commissions. See also Mulvey, 16].

2.1.4 Rebalancing an existing portfolio

It is perhaps not the most usual situation for a portfolio manager to have to construct a portfolio from scratch. More likely there's an existing portfolio (for instance last year's optimal portfolio) that needs to be re-optimized, or rebalanced. Of course one could just sell the old portfolio, then buy the new, optimal portfolio, but that would incur probihitive transactions costs, and would seem unnecessary if the changes in position of most instruments were small. It is therefore necessary to model transaction costs, especially the xed ones, explicitly to encourage the model to limit turnover. Using additional parameters, x0 to represent the existing i portfolio, this can (as in Section 2.1.3) be done using binary variables the precise model is left as an exercise to the reader (the model should take both xed and variable costs into account). In some situations certain holdings may be frozen, i.e., not to be changed. This can happen when the portfolio manager, for whatever reason, needs to keep her position in certain instruments. One can either \ x" the related variables at the current holdings or totally remove such holdings from the model by removing them from the universe U and modify the liabilities by subtracting the instruments' cash- ows. This is one example where the resulting liabilities may become negative.

2.1.5 Time-mismatched cash ows

We have implicitly assumed that all incoming cash ows in a given time period was available to meet the liabilities in that period. This may not always be the case. If the time periods are years, and the liabilities are due on the rst day of the year, then some of the cash ow from assets such as bonds would not be available yet. One solution is to augment the set of time periods T to include all cash ow and liability dates, but this leads to a larger model, and is usually not necessary. An alternative is to assume that money can be borrowed against certain future cash ows (still within the same time period), and that incoming 8

cash that arrives before it is needed will be invested until needed. In this case the cash ows can be adjusted in time value. For instance, if a bond coupon payment ci arrives 6 months after it is need to cover a liability, and the annual rate of borrowing against it is , then the value ci (1 + ) 1=2 (or ci e =2 ) should be used instead of ci in the model.
; ;

2.1.6 End-of-horizon e ects

Many types of liabilities do not have a speci c horizon beyone which they have been paid out. For instance, the life insurance company which insured Methusalem would need a longer than usual horizon to accurately model their liabilities. Similarly, some assets, such as stocks, do not have de nite horizons. However, the CFM models need a nite horizon. The usual way to handle this problem in practice is to calculate a nal liability (and a nal cash ow for the assets) at the model horizon, m, equal to the present value (at time m) of the remaining, outstanding liabilities or cash ows. This again introduces a risk to the model because an assumption needs to be made regarding future interest rates, or the future, present value of the outstanding cash ows. By using conservative estimates and a long horizon one can safeguard against surprises, but there is an inherent trade-o between model accuracy and complexity.

2.1.7 Diversi cation considerations

Many institutional investors have limitations on the allowable exposure to risky investments, such as corporate or municipal bonds, mortgage-derivatives, etc. Such limits are usually expressed as a maximum percentage of the portfolio value that may be invested in certain classes of investment vehicles. Such diversi cation, or limitation, constraints are easily modeled. Assume that an investor wishes to limit her exposure to the bonds in some set S U to a fraction, p, of total value, and assume that the instruments in question have positive cash ows once they have been purchased at a price Pi . The appropriate constraint can then be written as follows:

X
i S
2

Pi xi p

X
2

i U

Pi xi :

Naturally, there can be several such constraints covering di erent subsets of the asset universe. If a limitation expressed in face value amounts, rather than value is desired, the Pi s above can just be omitted.

2.1.8 Net present values and duals

To the optimal solution of an optimization program such as an LP is associated so-called dual variables. Dual variables complement the model (or primal) variables, but are associated with the model's constraints, including the bounds on variables. Expressed informally, the dual value corresponding to a constraint
linear expression

constant

tells us the change in objective value per unit change of the constraint's right-hand side, up to a point. If the constraint is tightened, the objective either stays the same (if the dual is zero) or gets worse (i.e., larger in a minimization program, smaller in a maximization program), and vice-versa if the constraint is loosened. For equality constraints the rules are slightly more complicated. The dual variables of the cash- ow constraints of CFM are interesting because they represent the present value of money at the time with which the constraint is associated. It is easy to see that this is correct: If the right-hand side of such a constraint (i.e., the liability) were increased by one unit, then the objective would increase by the additional cost needed today to cover the additional future liability. This notion of present value is not the usual one, obtained by discounting according to a zero-coupon yield curve. Rather, it is determined implicitly through the prices and cash ows of the assets in U . It is interesting to note that there is no explicit notion of \yield curve" in CFM, yet the model automatically \bootstraps" one.

2.2 A variation: Maximizing Horizon Return

incentive for the model to carry any surplus cash forward to the end, considering the possibility of \wasting" money by simultaneously reinvesting (at a low rate) and borrowing (at a high rate). We therefore consider two changes to the model: 1. the objective of the model is now to maximize the nal cash position, subject to meeting the liabilities, 2. Instead of minimizing an initial lump sum, it is assumed that there is a budget for investing. Hence, we now model an investor with a speci c sum to invest, with liabilities (possibly 0) to satisfy, who wants to maximize investment performance. The focus is now on investing rather than dedication. 10

CFM ignores what happens to the nal surplus cash, rm, and there is in fact no

Maximize Subject to
xhrb

Maximizing horizon position

MAX:

=0 =h 0 I2U Here, B is the investor's budget, and h, the horizon position, is the objective, explicitly set equal to the nal amount available. The liabilities can be interpreted as the investor's desired consumption along the way. Note that the interpretation of the dual prices is di erent from CFM (what?).

P F x +b +B = r0 i0 i 0 iP U Fi t xi + (1 + t ) rt 1 + bt = Lt + rt + (1 + t ) bt
2

i U bm rm xi ri bi
2

t 1

2.3 The Stochastic Case

Unfortunately, it is not always realistic in practice that the future data, even in a relatively simple model such as the Dedication/CFM or MAX model, are known with certainty. In fact, large nancial institutions (banks, insurance companies) often make a business precisely in taking on stochastic liabilities, thus freeing cheir customers from assuming risk. An example is annuities: The insured (annuitant) is relieved of the risk of outliving his funds, but the insurance company now has a stochastic liability (of course, the company does not assume this risk for free. For each individual annuity contract, it expects to gain). We start with the MAX model with reinvesting and borrowing, but with stochastic cash- ows, including liabilities. The model can be written: Maximize Subject to
xhrb

Maximizing horizon position under uncertainty ~ h

2

i U ~m b rm ~ xi ri ~i ~ b
2

P F x +b +B = r0 i0 i 0 iP U ~ ~ ~ Fi t xi + (1 + t) rt 1 + ~t = Lt + rt + (1 + t ) ~t ~ b b
;

=0 =~ h 0

where the ~ indicates stochastic parameters. For simplicity we have dropped the index set indications t 1 and i 2 U t 2 T s 2 S they correspond to the previous models. Note that the investment variables, xi , are not stochastic, since they need 11

to be determined up-front. We have assumed that rst-period data are known with certainty, but there's nothing inherently important about this. On the other hand, one now has the opportunity to also let and be stochastic. As this model stands, it is too general for our use, and it is not even clear ~ what it means to \Maximize h". If the stochastic parameters (we avoid the term \random variables" to avoid confusion with the optimization variables) have general, continuous distributions, the problem is analytically intractable, except possibly if the dimensions involved are very small. The practical way to proceed is to assume that all distributions involved are discrete. This is justi ed in that, in principle, any distribution can be approximated arbitrarily closely by a discrete distribution. Actually, it is customary to assume that the random data are given by a set of scenarios, where a scenario s 2 S is given by a collection of (joint) realizations of the random data that is we are given data (Fist Ls h)t T i U t for each scenario s 2 S , and then have to determine optimal x h, rs and bs variables. We also assume that scenario probabilities, ps , are given, although these are not yet used in our model. Note that the variables rs , bs and h are stochastic in the sense that they are scenario-dependent, but they are still outputs from the model, not inputs. They are example of stochastic optimization variables, which we'll encounter again shortly. With a change of notation to emphasize the scenario-based nature of the model, we have:
2 2

Maximize Subject to
xhrb

P p hs s S s P F x +b +B = r0 i0 i 0 iP U Fist xi + (1 + t) rts 1 + bs = Ls + rts + (1 + t ) bs t t t

2 2

Maximizing expected horizon return

S-MAX:

i U bs m s rm xi ris bs i
2

=0 = hs 0

where we have made precise also the objective: Maximize expected horizon return. This is only one possible objective, corresponding to a risk-neutral investor we will see other possibilities later (max worst case, max expected utility etc.). For a practical application of a similar model see Adamidou et al. 1]. For a related approach, Scenario Immunization, see Dembo 8]. 12

It is clear that the set of scenarios chosen for S-MAX is extremely important for the optimal solution returned by the optimization. The scenario set needs to be: Comprehensive: It should capture all aspects, both extreme and \normal" instances, of the underlying distributions, Consistent: It must capture correlations among the stochastic data well. Comprehensiveness is important because extreme scenarios must obviously be present in the set so the model can \see" them and take them into account. On the other hand, the model should not be blindsided by extreme events (which are usually very low-probability events), so \normal" cases need to be present as well. This may not be too serious for S-MAX, where a low-probability event may not in uence the objective too much, but it could have a very large e ect on the portfolio generated by S-CFM. Consistency is equally important. In real life events are very often correlated. In Finance, some instruments pay out precisely when other instruments do not (puts and calls). If a risk-averse investor can nd negatively correlated assets, or assets which are positively correlated with the liabilities, they should be included in the model, so that their cash ows re ect this correlation. This means that the FIs t are consistent by not being uniformly \all high" or \all low". The scenario set must also be consistent in the time dimension. Some nancial instruments pay out a quite predictable amount, but the timing of payouts are unpredictable. Examples include CMOs or life-insurance. Hence, low initial payouts should correspond to higher, later payouts for the same instrument under a given scenario. We might want to add as a third requirement to the asset set that it should be small! The model above, with a reasonable time horizon and a reasonable number of assets is already big. If on top the size of the model is multiplied by j S j, it quickly becomes unsolvable. For this reason, Scenario Generation, or the way in which scenarios are produced to satisfy all three requirements, is in practice a major component of designing a stochastic program (or is simply ignored!). We give below some examples how scenario generation in Financial planning has been approached. Other examples from real life can be found in, e.g., Cari~o, Myers n and Ziemba 27], and Mulvey..Scenario Generation for the Towers Perrin ref..

2.4 Scenario Generation

Scenario-Generation in Fixed-Income Planning. Very often one can identify a single or a few underlying, stochastic processes, whose behavior can be modeled and which determines the stochastic parameters. An important example is nancial planning involving xed-income instruments, where the cash- ows of the instruments are (primarily) driven by the future development of interest rates.
13

The crucial point is that future cash ows must be unique functions of future interest rates, which is the case for Treasury Bonds, and to a large extend to many other xed-income instruments. This is one way to proceed: 1. One starts with an interest model, either a one-factor like Vasicek which models short rates (see Hull 11], 17.4), or a multi-factor like Nielsen-Ronn 23], which models two factors, short and long rates. The model should be calibrated so it is consistent with expectations of future events. This can be done by matching the model to historical observations (such as the volatility and correlations of the factors), or by matching the model to observable prices of traded instruments (mostly options), or by the decision maker simpley postulating a set of interest rate scenarios (s)he feels covers the possibilities. 2. The interest rate model, if not already in (discrete) scenario form, is then used to generate scenarios of future interest rates. The scenarios need to indicate the whole term structure at each future time point in the model. In the case of a popular model such as Vasicek, this step could involve simulation. 3. Finally, given these scenarios, the xed-income instruments' cash ows are calculated. This may often be done only approximately. For instance, with corporate bonds or mortgage-securities, interest rates are only one determing factor. The company may go bankrupt for un-economic reasons, or the home-owner get divorced and sell the house. Examples where this approach was taken, and where the liability side was also stochastically dependent on interest rates is Nielsen-Zenios 25], upon which the Winvest case is built upon.

The New-York 7 The so-called \NY-7" is a set of interest rate scenarios which

is used by the industry regulators in New York. It is a set of 7 interest rate scenarios, and each nancial intermediary in the state of New York each year needs to demonstrate to regulators that they, with their present asset and liability portfolios, would stay solvent if interest rates over some speci ed future horizon behaved according to a NY-7 scenario. Since many assets and liabilities are not governed by interest rates alone, this is a very inexact science, but it actually forces nancial institutions to look hard at their asset-liability matches and validate (internally and externally) their approaches in calculating future in- and outows. The NY-7 only considers parallel shifts to the term structure. Starting with today's term structure, they are: 14

1. Rates stay the same, 2. Rates increase by 50bp per year for 10 years, 3. Rates increase by 100bp per year for 5 years, and then drop by 100bp per year for 5 years, 4. Rates increase by 300pb, then stay the same, 5. Rates decrease by 50bp per year for 10 years, 6. Rates decrease by 100bp per year for 5 years, and then increase by 100bp per year for 5 years, 7. Rates decrease by 300pb, then stay the same, A \bp", or basis-point, is 1/100 of a percentage point, so an increase of 50bp of an interest rate of, say, 5.5% would take it to 6.0%. All rates are truncated at 0%. The scenarios are equally likely. The NY-7 has a time horizon of 10 years, and includes both high, immediate shifts, long, slow shifts, and reversing shifts. Since the changes are quite dramatic, it is often argued (and hoped) that a model which hedges well against the NY-7 will be well-hedged against virtually any real change in rates, possibly with the exception of non-parallel shifts. Many institutions are content with using just the NY-7, possibly adding a few more scenarios, for planning purposes. While the NY-7 certainly contains dramatic (extreme) events, it can be argued that it probably leaves companies over-hedged. (Under networks cite 18] perhaps 17] under stochastic networks cite 20]). Under diversi cation/ xed income 22].

3 The Two-Stage Stochastic Program

The models in the previous sections were static, meaning that the initial decision (portfolio) is never modi ed. We now introduce dynamic models, where future changes, recourse, to the initial decision are allowed. This leads naturally to Twoand Multi-stage Stochastic Programs. Although the stochastic cash- ow matching problem and the S-MAX model for maximizing the expected horizon return are better approximations to an uncertain world than their deterministic counterparts, they have a serious limitation: They model a strict buy-and-hold situation, where the investor buys a portfolio, then 15

3.1 Modeling future rebalancing: Dynamic models

never changes it. There is, of course, nothing to keep the investor from periodically reviewing the portfolio, but then we might as well model this possibility explicitly. We rst consider a simpli ed model where the decision maker can rebalance once in the future, after one time period. To do this, we introduce, in addition to the initial, or rst-stage, decision variables xi , a set of second-stage variable, yis, where yis is the holdings of security i 2 U in time periods t = 1 ::: m, under scenario s 2 S . Note that the second-stage variables are scenario-dependent. This is natural, otherwise we would force the same second-stage portfolio no matter what was learned about the economy (or cash ows) during the rst period. The S-MAX model is modi ed as follows: Maximize Subject to
xyhrb

=0 = hs 0 There are two new things in this model: First, the x and y variable now explicitly relate to di erent time periods, or stages in the decision process. The rst stage is at time t = 0 where the xi are determined, and the second stage is at time t = 1 where the yis are determined. Second, we need constraints that link the second-stage to the rst-stage variables, since the portfolio cannot change value (possibly except for rebalancing transaction costs, which are ignored here). This is expressed by the third constraint above, stating that the total value, at time 1, of the existing portfolio, sumi U Pisxi , must equal the total value of the new P portfolio, i U Pis yis, and where Pis is the price of instrument i at time 1 under scenario s. We have now introduced new, stochastic parameters to the model, namely the future prices. This means that we need a way to estimate these prices. A reasonable way to do this is very problem-dependent. A possibility in a xed-income world is to let the price equal the present value (at time 1) of the future cash ows, depending upon scenario, but this requires some assumption (or knowledge) of future interest rates. If the model is generated from an interest-rate model, such knowledge may be available. If the rst-stage variables are restricted to shortmaturity bonds, then the price-uncertainty is smaller, and the dynamics of the
2 2

P p hs s S s P F x +b +B = r0 i0 i 0 iP U Fist yis + (1 + t ) rts 1 + bs = Ls + rts + (1 + t ) bs t t t i U P P sx P P s ys =

2 2 ; 2

A Two-Stage Stochastic Program

2S-SP:

bs m s rm xi yis ris bs i
2

i U i i

i U i
2

16

model come from the price changes in the longer, second-stage universe. There seems to be no general way to answer, or get around, this question. Exercise: How could trading costs be incorporated into the above model?

3.2 A Formalization of the Stochastic Program

The two-stage stochastic program is a problem class of such importance that it deserves to be viewed in a more general and formal framework than that of nancial planning. What follows is a general introduction to SPs, without any speci c application in mind. The exposition here is from Nielsen and Zenios 26]. The two-stage stochastic linear programming model addresses the following situation: A decision is made at the present time facing future uncertainties. At a future time the uncertainties are resolved, and a recourse action is taken. The uncertainties of the model are represented by stochastic parameters, that is parameters whose actual values are not known until a future period, but their probability distributions are known a priory. The decision to be made at present is called the rst-stage decision, and the (future) recourse decision is called the second-stage decision, which is contingent upon the rst-stage decision and on the observed realization of the uncertain parameters. In algebraic notation, the problem can be formulated as follows: SLP] Minimize cT x + Q(x) (1) n
x
2<

Subject to Ax = b 0 x u: where

(2) (3)

E is the expectation operator de ned on some probability space ( F P ), and Q(d h T v W j x) = Minimize dT y (4) n
y

Q(x) = E fQ(d h T v W j x)g

2<

Subject to Wy = h ; Tx 0 y v:

(5) (6)

Bold letters are used to designate stochastic quantities, and the corresponding roman letters designate instances of the stochastic quantities. If the second minimization problem is infeasible, we take its value to be +1. In this paper we 17

assume that the problem has relatively complete recourse, i.e., problem (4) { (6) has a feasible solution for any value of the rst-stage variable x which satis es (2)- -(3). In this formulation, n1 and n2 are the number of rst-stage and second-stage decision variables, respectively. There are m1 rst-stage constraints (2), and m2 second-stage constraints (5). All other matrices and vectors have conformable dimensions. Uncertainty of the second-stage problem is represented by the stochastic quantities d, h, T, v and W. In this paper we consider the case where the stochastic quantities have a discrete and nite joint distribution, represented by the scenario set = f1 2 3 : : : S g. This set can be an exhaustive enumeration of the possible realizations of the stochastic parameters, or it can be obtained by sampling from the joint distribution, see, e.g., Dantzig and Infanger 1991]. In this case, we have

Q(x) =

S X

where ps is the probability of realization of scenario s, ps = P f(d h T v W) = (ds hs Ts vs Ws )g for all s 2 : (8) PS p = 1. It is assumed that ps > 0 for all s 2 and that s=1 s Assuming that the uncertainties are represented by a nite scenario set, the stochastic program SLP] can be reformulated (Wets 1974]) as the following deterministic equivalent program:

s=1

ps Q(ds hs Ts vs Ws j x)

(7)

DELP]

Minimize
x
2<

cT x +

S X

Subject to Ax

n1 ys 2<n2

s=1

psdT ys s b hs for all s 2 , u, vs , for all s 2 .

(9) (10) (11) (12) (13)

= Ts x + Ws ys = 0 x 0 ys

This large-scale, linear problem consists of n1 + S n2 variables and m1 + S m2 equality constraints. The constraint equations can be written in matrix form as the following dual block-angular system:

0 B B B B B B @

A T1 W1 T2 W2
.. .

TS

...

WS
18

1 C C C C C C A

0 B B B B B B @

x y1 y2

yS

1 0 C B C B C=B C B .. C B A @ . C B

b h1 h2

hS

1 C C C C .. C A . C

(14)

It is now clear that, if a vector x is given, satisfying (2){(3) for the rst-stage ^ variables, the system for the second-stage variables can be written as:

yS hS TS This system can be decomposed into S scenario blocks and solved in paralle l.

0 B W1 W2 B B ... B @

WS

1 C C C C A

0 B B B B @

y1 h1 y2 C B h2 C C B C C B C

0 B B ^ B .. C = B .. C ; x B . A @ . A @

1 0

T1 T2 C C C
.. C . A

(15)

This observation is the basis for the frequent use of decomposition algorithms in solving large-scale SPs. For a short introduction to Benders decomposition in this context, see Nielsen and Zenios 26]. Exercise: Show in detail how the various matrices and vectors in the formulation in this section relates to the formulation 2S-SP in Section 3.1.

3.3 A Network Formulation

We now return to nancial planning model very similar to 2S-SP from Section 3.1. The purpose of this section is to show the power of networks in visualizing an optimization problem, and the almost mechanical way in which an algebraic formulation of the problem can be arrived at, once the graphical network is correctly understood. De nition: A Network is a directed graph... ( ll in later). The problem addressed in the following is very similar to the Winvest case. It is a two-stage stochastic network, where the rst-stage decision is an investment, and the second-stage decision is a subsequent rebalancing. There is one major di erence, however, between the following formulation and 2S-SP: The decision variables represent the actual value of holdings, not the face value. So a holding of 5 bonds, each with a face value of $100 and a price of $92,is represented by some xs = 460, not $500. It turns out this change in decision variables is very pi convenient for network formulations. The paper by Nielsen and Zenios 25], from which the material in this section is adapted, funds an insurance company's exposure from selling an annuity product called a Single Premium Deferred Annuity, or SPDA. The portfolio is to be constructed from a universe U of nancial instruments. The model is dynamic, and all events, such as asset trading or coupon payments, occur at discrete time points, t = 0 ::: T . An initial portfolio is constructed at time 0 = 0, and is subsequently rebalanced at time points 1 < 2 < ::: < Y = T . During the periods between rebalancing points (where period p is the time span p 1 t < p ) the portfolio composition remains unchanged, except that cashows are reinvested at the short rate. Transaction costs are included, and limited borrowing is allowed.
;

19

Uncertainty is modeled using a set of interest rate scenarios, generated according to a suitable term structure model. Lapse behavior for SPDA annuitants is driven by the short-term rates under each scenario, leading to a stochastic liability stream. On the asset side, the prices of assets in future time periods, as well as the short-term borrowing and lending rates, are also interest rate driven. Their estimation is described in Section 4.2 in 24]. The objective of the stochastic model is to construct a portfolio whose cash- ows match the liabilities in each time period under all scenarios, while having a risk-return pro le consistent with a prescribed level of risk tolerance. In its basic form the model is a two-stage, stochastic program with recourse. The rst-stage decision is the construction of the initial portfolio. After a realization of interest rates has been observed, the portfolio is rebalanced. The construction of the rebalanced portfolio constitutes the second-stage decision. Of course, rebalancing decisions are contingent upon the realized scenario and the composition of the initial portfolio. For a given interest rate scenario, the model has a network structure, as shown in Figure 1 for two assets (Mulvey and Vladimirou 21]). Columns of nodes correspond to di erent time points, while rows of nodes correspond to di erent assets. The bottom row of nodes corresponds to cash. Horizontal arcs | between nodes corresponding to the same instrument i 2 U | model the holding of instrument i in the portfolio. Arcs which link nodes in the bottom row model short-term cash reinvestment and borrowing. The vertical arcs which link cash and instrument nodes model changes in the position of each instrument, that is, rebalancing. The initial infusion of cash into the model consists of the SPDA premium and possibly insurer equity, and is used for constructing the initial portfolio. Triangles on arcs designate arc multipliers, i.e., that the value (cash or holding) entering the arc is changed by some proportion (multiplier) before leaving the arc. For arcs modeling sales of assets, this multipler represents a transaction cost on sales. For reinvestment and borrowing, the multiplier represents short-term reinvestment and borrowing rates. Multipliers on the arcs which represent the presence of instruments (holding arcs) model the yield of the instrument for each time period. Of course, these multipliers are dependent on the speci c scenario, S = f1 ::: S g. The arcs in bold type for investments at time 0, i.e., construction of the initial portfolio, are rst-stage decisions, and must be the same regardless of the scenario. The remaining arcs (later period holdings and rebalancings) represent second-stage decisions which are allowed to depend on the scenario realized. This distinction between rst- and second-stage decisions is a key feature of two-stage, stochastic models, where immediate decisions cannot depend on, as yet, unrealized data, but future decisions can. 20

Period 1

Period 2

... t=p

Period Y

t=0 s x1i
Instrument 1

t=1

s mpi

t=T

s xpi

y0i

s y1i

s z1i

s ypi

s zpi

s zYi

Instrument 2

1-

Cash

vs 1
s u1

vs 2 us 2

L1

L2

LY

First Stage Variables (Scenario independent) Second Stage Variables (Scenario dependent)

Figure 1: Network model underlying the two-stage, stochastic SPDA model. This gure includes two instruments, and depicts a 4-period model. Stochastic quantities are denoted by a superscript, s.

21

We now describe the components of the stochastic model. The model is described in three parts, dealing with security holdings, cash position, and the model's objective. The complete notation and algebraic formulation of the model is given in Appendix A. value) of instrument i during period p under scenario s. Purchases and sales of s s securities are represented by the variables ypi and zpi , respectively, where the index p refers to transactions which take place at the end of period p. The constraints which de ne the holdings of securities during each time period under each scenario can be written as:

3.3.1 Model De nition

Security Holdings: The variables xs are used to represent the holding (in dollar pi

xsp+1)i = (

xsi = 1

y0i s xs ; z s + ys mpi pi pi pi

for all s 2 S i 2 U : (16) for all s 2 S i 2 U and p 2 1 ::: Y ; 1: (17)

Equation (16) states that the initial holdings equal the initial investments. The initial investment, y0i , must not depend on the scenario to be realized, and hence does not have the scenario superscript. These variables are all measured in cash value rather than face value. Equation (17) de nes the changes in the cash value of holdings due to yields, sales and purchases of instruments. The multipliers ms pi represent the yield during period p, of instrument i, under scenario s. We assume that the nal portfolio must be liquidated, and this is ensured by the following constraint, which states that sales after the last period must equal holdings after that period: s zY i = ms ixs i for all s 2 S and i 2 U : (18) Y Y

Cash Position Accounting: Next come de nitions of cash positions in each time period. The initial amount of cash available (premium and equity) is denoted by C . Cash is used for purchases of securities and the payment of liabilities, and is generated by sales. During each period, excess cash is invested at the short rate, and borrowing is allowed. We use us to designate short-term cash p s investments during each time period, and for each scenario, and vp to denote short-term borrowing. The constraint for the rst time period states that the initial amount of cash s available, C , plus any rst-period loan, v1 , must equal the amount invested in instruments, y0i or held in cash during the rst period, u1 :
X
2U

s y0i + us ; v1 = C for all s 2 S : 1

(19)

22

In general (although not for the rst time period), short-term borrowing and cash investment are scenario-dependent, hence the superscript s. The corresponding cash- ow balance constraints for the intermediate time periods are:

X
2U

s s s s s s ((1 ; )zpi ; ypi) ; us+1 + (1 + rp)us ; (1 + rp + )vp + vp+1 = Ls p p p

(20)

for all s 2 S p = 1 ::: Y ; 1

where is a transaction cost (charged as a fraction of the amount of the transaction) when selling instruments and Ls is the liability due at the end of period p p under scenario s. For simplicity, transaction costs on purchases are not used. Rather, should account for total costs. The cash- ow constraint for the last time period is:

X
2U

s s s (1 ; )zY i + (1 + rY )us ; (1 + rY + )vY = Ls + W s for all s 2 S : (21) Y Y

It di ers from (20) in not allowing purchases, and in de ning the nal wealth under each scenario, W s, as the surplus cash after the last liability, Ls , has been Y paid.

Objective Function: The objective of the model is to maximize the expected value of a measure of nal wealth across scenarios. We maximize the Expected Utility of Return on Equity. Return on Equity, ROE, under scenario s is de ned as rs = W s=E , where E is equity. Utility is measured using the family of iso-elastic utility functions (Ingersoll 1987] 13]):
( 1 U (r) = log(rr )
1
;

1;

; 1) for 6= 1
for = 1

(22)

where 0 is a risk-aversion parameter. Higher values of implies more riskaversion, i.e., less tolerance for risk on the investors' part. The value = 0 results in a linear utility function which corresponds to a risk-neutral attitude, and = 1 in a logarithmic utility function which corresponds to a moderate level of risk-aversion and which is known as the growth-optimal strategy. The properties of logarithmic utility functions for portfolio selection are discussed in McLean, Ziemba and Blazenko 15]. Assuming that all scenarios are equally probable, the objective function of the model is de ned by X : Maximize Expected Utility = u = 1 U (W s=E ) (23)

jSjs

2S

where u is the expected utility of return on equity. 23

Utility measures in themselves are not meaningful, except for ranking uncertain outcomes. Hence, we employ the more meaningful certainty-equivalent return on equity, CEROE. The certainty-equivalent has the same utility as the expected utility of the investment. The investor with the prescribed risk-attitude is indifferent between receiving the (deterministic) certainty-equivalent return and the (stochastic) portfolio return. We use the CEROE value of di erent portfolios when comparing alternative portfolio management strategies. The Certainty-Equivalent is de ned by CEROE = U 1 (u ) (24) and is used in comparing investors' preferences among di erent investments. A complete description of the model is given in .
;

3.3.2 Extension to a Multistage Model

The two-stage model described above allows rebalancing decisions at times t > 0 to depend on future prices and returns, i.e., on data not yet known at time t. This implies that the decision maker has \perfect foresight" in making rebalancing decisions after the rst time period. This is of course unrealistic. To overcome this problem, the two-stage model is extended to a multistage model in which decisions at time t > 0 do not depend on the speci c sequence of events which will be realized during later time periods, but depend only on events observed prior to time t. Multistage models are a much better representation of reality than two-stage models. They have the same data requirements as two-stage models, but are substantially more complex in structure, and can be signi cantly harder to solve. Similarly to the two-stage models, they are based on scenarios of future interest rates. In the two-stage case these scenarios are independent of each other but in the multistage case, they are grouped together such that certain scenarios are indistinguishable from each other up to a certain time point (See Figure 5 in 24]), where 8 of the 16 interest rate scenarios are indistinguishable up to time 6, groups of 4 are indistinguishable up to time 12, etc.). The lack of foresight mandates that decisions made under such indistinguishable scenarios, up to the time point where they di er, should be the same under each scenario. This leads to the requirement that some variables should be identical across scenarios. We return to multistage models in Section 4.

4 Multistage Stochastic Programs

As alluded to in Section 3.3.2, the two-stage stochastic program can be extended in a natural fashion to a multistage program. This program models the situation 24

where there are several opportunities for adjusting ones position, such as rebalancing a portfolio. Each such opportunity is again called a stage. The following is a formal introduction to the problem (from Nielsen and Zenios 24]). We now de ne the linear, multistage stochastic problem with generalized network recourse, followed by a deterministic equivalent formulation. Transposition of a vector x is denoted xT , and the inner product xT y is written xy when context makes the meaning clear. Bold letters are used to denote stochastic quantities, and the corresponding roman letters designate instances of these quantities.

4.1 Formulation of the T -stage Stochastic Program

n h MS] min c1 x1 + E 2 min c2 x2 + E 3
s.t.
x1

A T -stage stochastic programming problem can be formulated as follows (Birge 2]):

A1 x1 B2x1 +A2x2 B3 x2 + A3x3

x2

min c3 x3 +
x3

+ E T 2 :::
j

T ;1 min cT xT

io

0 xt ut for t = 1 ::: T where

t = (At

= = = .. ... . BT xT 1 + AT xT =
;

b1

xT

b2 b3 bT

are random variables, i.e., Ft -measurable functions t : t 7! <Mt on some probability spaces ( t Ft Pt ). The decision variables xt 2 <nt , for t = 2 ::: T , are stochastic variables measurable on the - eld generated by t . The notation E denotes mathematical expectation with respect to , and E i j similarly denotes conditional expectation.
j

Bt bt ct ) for t = 2 ::: T

4.2 Scenarios and the Scenario Tree

The sequential nature of the decision process is apparent from this formulation. At each stage of the decision process, a conditional expectation is to be minimized. We consider in this paper the case where each t can assume only nitely many values. Then a scenario tree, as in Figure 2, can be used to represent the way in which the stochastic variables t evolve: The root of the tree corresponds to the immediately observable, deterministic data, A1 B1 b1 and c1 . The nodes 25

Stage 1:

Stage 2:

= B B BBN

ZZ ZZ

t=1

ZZ ~

Stage 3:

B B BBN

B B BBN

t=2

t=3=T

Figure 2: Scenario tree for a 3-stage program (T = 3) having 6 scenarios. In this example, 2 has three possible realizations. 3 has two possible realizations for each realization of 2 . of the tree at level t 2 corresponds to possible realizations of t. These set of possible realizations (or, equivalently, their probabilities Pt ) at level t are, in general, dependent on the preceding observations of 2 ::: t 1 . The nodes in the scenario tree are also associated with the sequential decision process, xt , so that each node at level t corresponds to a decision xt which must be determined at time t, and which could depend on 2 ::: t as well as x1 ::: xt 1 . This process must, of course, be adapted to t in the sense that xt cannot depend on the speci c future events, t+1 ::: T which are not yet realized. In the formulation MS], the constraints on xt are in uenced directly only by xt 1 , not by xt 2 ::: x1 . This structure is not a loss of generality (by suitable choice of the state space), and is not imposed by our algorithm, but is used because it conforms more closely to the scenario tree, where quantities associated with a node at level t depends directly only on quantities associated with the predecessor node of the tree. The ancestor node is for a node v represented algebraically by the ancestor mapping, a(v). We need to introduce the notion of \scenario", and its relationship to the scenario tree. A scenario is de ned as a possible realization of the stochastic variables 2 ::: T . Hence, the set of scenarios, = f1 ::: S g is in a one-toone correspondence with the set of leaves of the scenario tree, and we associate scenario s 2 with the sth leaf of the scenario tree. With this correspondence, we use the notation a(s) s 2 , to denote the ancestor of the sth leaf node. The probability of scenario s is denoted by ps > 0.
; ; ; ;

26

4.3 The Split-Variable Formulation

We introduce now the split-variable formulation of MS], see, e.g., Rockafellar and Wets 28] or Ruszczynski 29]. The split-variable formulation associates with each scenario s 2 and each stage t = 1 ::: T a set of decision variables, xs 2 <nt , t which replace the stochastic variables xt used in MS]. For each scenario s 2 , de ne the (deterministic) scenario subproblem:

SUB(s)]

min s s.t.

xt t=1 ::: T

cs xs + cs xs + 1 1 2 2

+ cs xs T T = = = .. ... . s BT xs 1 + As xs = T T T
;

A1 xs 1 s B2 xs +As xs 1 2 2 s B3 xs + As xs 2 3 3

b1 bs 2 bs (25) 3 bs T

0 xs ut for t = 1 ::: T t

where superscript s on the data, As Bts bs and cs , denote the unique realization t t t of t associated with scenario s. The scenario subproblems are obviously independent of each other. In order to obtain a formulation equivalent to MS], we need to impose further constraints linking together the scenario subproblems. The non-anticipativity constraints impose the logical requirement that decisions up to stage t must coincide for those scenarios which have common root-to-leaf paths up to level t in the scenario tree. Letting aT t denote the (T ; t)th power of the mapping a, the non-anticipativity constraints can be written:
;

xs = xs+1 for t = 1 ::: T and s 2 : aT t(s) = aT t (s + 1): t t

; ;

(26)

The complete set of scenario subproblems, together with the non-anticipativity constraints, de ne a program equivalent to MS]. The split-variable formulation is:

SPLIT] min
xs t=1 ::: T t

X
2

ps (cs xs + cs xs + 1 1 2 2

+ cs xs ) T T = = = .. ... . s BT xs 1 + As xs = T T T
;

s.t.

A1 xs 1 s B2 xs +As xs 1 2 2 s B3 xs + As xs 2 3 3

b1 s 2 bs s 2 2 bs s 2 3 bs s 2 T

(27)

27

Stage 1:

x1 1 x1 2 x1 3

x2 1 x2 2 x2 3

x3 1 x3 2 x3 3

x4 1 x4 2 x4 3

x5 1 x5 2 x5 3

x6 1 x6 2 x6 3

Stage 2:

? ?

? ?

? ?

? ?

? ?

? ?

Stage 3:

Figure 3: Illustration of the split-variable formulation corresponding to the scenario tree of Figure 1. Variables associated with each scenario/stage combination are shown. Horizontal double-lines indicate non-anticipativity (equality) constraints among these variables. 0 xs ut for t = 1 ::: T and s 2 t xs = xs+1 for t = 1 ::: T and s 2 : aT t(s) = aT t (s + 1): t t
; ;

The split-variable formulation is attractive for any algorithmic framework where the non-anticipativity constraints can be temporarily ignored while the S independent subproblems are solved, possibly simultaneously and in parallel. This is the case with the progressive hedging algorithm of Rockafellar and Wets 28] as well as with the augmented Lagrangian decomposition method of Ruszczynski 30], the diagonal quadratic approximation algorithm of Mulvey and Ruszczynski 19] and the proximal minimization algorithm we study here. (((For more discussion of the structure and formulation of multistage models, see Gassmann 10] or Nielsen and Zenios 24]. For a state-of-the art overview see Birge 3]. Infanger (1994) 12] is a monograph on the use of importance sampling within Benders Decomposition.)))

5 Immunization: Present Value, Duration, Convexity

The concepts of \Duration" and \Convexity" have a long history in asset-liability management, dating back to Macaulay 14], who in 1938 proposed duration as a measure of the sensitivity of a bond portfolio's value to changes in interest rates. We consider again a company (such as an insurance company) which has a future liability stream, and which wants to fund these by a bond (or some other xed-income) portfolio. We look for an asset portfolio which has su cient value to fund the liabilities, but we relax the requirement of cash- ow matching. Instead, 28

we require that the sensitivity to changes in interest rates is equal on the asset and liability sides. That is, we immunize against interest rate changes. Many insurance (and other) companies got into trouble in the late 1970's and early- to mid-eighties where interest rates (in the U.S) rose. Consider the following example. An insurance company in 1985 sold a GIC (Guaranteed Investment Contract), worth $1,000,000, maturing in 1992 at a rate of 5%. That is, in return for the investment, the insurance company promises to pay $1 000 000 1:057 = $1 407 100 in 1992, 7 years later. To ensure this liability, the insurance company (in this simple example) purchased the highest-yielding zero-coupon they could nd, a 30-year, yielding 7%. The face value of this investment was 1 000 000 1:0730 = 7 612 300. In 1992, the general level of interest rates had risen by 2%, so the (now 23year) zero now had a yield of 9%, and hence the company's investment was worth 7 612 300 1:09 23 = 1 048 800, | 358,300 less than the liability! A very modest change in interest rates lead to a large shortfall, or asset-liability mismatch. The problem here was that the assets and liabilities were duration-mismatched, i.e., reacted very di erently to changes in interest rates. This problem is precisely what immunization attacks.
;

5.1 Present Value and Dollar Duration.

The key formulas we will be using tie together the future cash ows of a bond with its present value, or price. It's important to note that when talking about duration, an underlying assumption is that the yield curve, whatever its shape, only moves in \parallel" up or down. More advanced techniques (bucket or factor immunization) address the general case of yield curve shape changes. By \yield curve" we refer to the annualized yield of a zero-coupon, default-free (e.g., Treasury) bond with given maturity t, called yt . A simplifying assumption would be a at term structure, where all yt are identical we will assume a possibly non- at term structure. Let Fit be the cash ow (coupon and/or principal) of a bond i at time t, and let T be the set of time points where cash ows occur, measured from time 0 (\now"). Then the present value of the bond's cash ows is given by

Pi =

X
2

t T

Fit (1 + yt ) t :
;

(28)

Ideally, this value is also the price of the bond (although small deviations in richness/cheapness are common) - otherwise there would be arbitrage opportunities between the par and zero coupon bonds. Notice that cash ows that arrive at di erent times are discounted by di erent factors. The yield to maturity of a bond is the \average" yield | constant through 29

time | which would equate the right-hand-side of the expression above to the bond's price, i.e. the yield ri which solves:

Pi =

X
2

t T

Fit (1 + ri )

(29)

with Pi given in (28), or being the observed market price. We will be interested in protecting the value of a (bond) portfolio against changes in interest rates. As mentioned, we assume that interest rates only change \in parallel", i.e., that the whole yield curve shifts up or down by the same amount. How does a bond's value change in response to such changes? For simplicity we write the bond's price as a function of its yield-to-maturity (YTM), and then ask how sensivite the price is to changes in YTM. This is equivalent to assuming a at term structure with constant discount rate, ri , and considering parallel shifts up or down. A natural measure of price-sensitivity is dollar duration, which is just the derivative of the price with respect to interest rates: X Didol = dPi = ;t Fit (1 + ri ) (t+1) : (30) dr
;

Notice that dollar duration is negative: If interest rates rise, present value drops. Zenios et al 6] use ki for dollar duration.

t T
2

5.2 The Immunization Model

Let PL be the present value of a future liability stream (calculated using the same expression as for Pi ). We would like to construct a portfolio of bonds (from a bond universe U ) such that the portfolio's present value matches that of the liability stream: X Pi xi = PL where xi is the face value holdings of bond i. In addition, we would like to immunize the portfolio against small, parallel shifts in interest rates | in other words, make sure that for small shifts, the value of the portfolio remains nearly equal to that of the liabilities. Dollar duration matching achieves precisely this:
i
2U

X
2U

Didol x = kL

The classical immunization model (e.g., Zenios et al 6]) is an LP with the above two constraints, plus non-negativity conditions on the xi . It is tempting to seek a cheapest portfolio which satis es this, but that is generally meaningless, since the price of the portfolio, by the rst constraint, is already xed to the 30

present value of the liability. To the extend that the present value and the price of the obtained portfolio di er, the model is (in a fair market) just picking mispriced bonds. Another possibility is to seek to maximize the portfolio yield, which can be approximated as: X Maximize ; Didol ri xi : This expression is explained in 6], but it is useful to think about it as maximizing the yield of the portfolio (given by the bond YTMs, ri ) weighted by the duration of the bond (Didol ), i.e., how long the bond's yield a ects the portfolio yield (see below for the connection between dollar duration and \average time" till we receive the bond cash ow). This objective is subject to the same objection as minimizing the portfolio price, since low price (possibly due to mis-pricing) is of course associated with appearant higher yield.
(Simple Immunization Model) Maximize ; P Ddol r x
x U
2

2U

Immun-1:
i

Subject to

P P x Pi U Didol i x
2 2 2

i U i

xi

i U i

= PL dol = DL 0 i2U

5.3 Macaulay and Modi ed Duration

Dollar duration is not the only measure of price sensitivity used. The traditional, and more intuitive, measure of \duration" is Macaulay Duration (DiMac ). Dollar duration is an absolute measure: If portfolios A and B have identically timed cash ows, but A's cash ows are twice as big as B's, then A has twice the dollar duration of B (and both are negative). It is often preferred to used a relative duration measure, where the timing and relative magnitude of the cash ows is what matter, not their scaling. This introduces Modi ed Duration, X Dmod = ;Ddol =P = 1 t F (1 + r ) (t+1) (31)
i i i

Pi

t T
2

it

Note that portfolios A and B have the same modi ed duration. Modi ed duration is easy to use in estimating price changes when interest rates change. Consider (from Tuckman) a bond with Pi = $100, a yield ri = 8% and a modi ed duration of Dimod = 5. To estimate the change in price when rates increase by, say, 10bp, just multiply the rate change by modi ed duration: Pi =Pi 5 0:001 which equals 0:5%. Hence the price falls by (approximately) 0.5% to $99.50. 31

Modi ed duration is measured in years, but unfornately, a single payment occurring at time t does not have modi ed duration t. Macaulay-duration has that property. It is de ned by multiplying modi ed duration by 1 + ri : X ;t F (1 + r ) t) (32) DMac = (1 + r ) Dmod = 1
i i i

Pi
;

t T
2

it

Macaulay-duration can be viewed as a weighed average of the times to arrival of 1 cash ows, where the weights Pi Fit (1+ ri ) t are proportional to the magnitude of the cash ow, and sum to 1 (why?). In estimating price changes using Macaulayduration, one multiplies DiMac not by the absolute interest rate change, but by the relative change (in 1+ ri ). The bond given above has DiMac = (1+ ri ) Dimod = 5:4 years, so if rates increase by 10bp from 8%, the price change is approximately 0:1%=(1 + 8%) DiMac = $0:50 again. Macaulay-duration is a more intuitive concept than modi ed duration, but for our purposes any of the three measures are equally good in hedging interest-rate uncertainty.

5.4 Critique of the simple Immunization Model

The immunization model Immun-1 has only two constraints, apart from nonnegativity. This means that in general, only two bonds will be chosen (only two xi will be positive). Typically the optimal portfolio is a \bar-bell" portfolio consisting of a very long bond (because they generally have the highest yield, hence contribute the most to the objective), and a very short bond (to match the liabilities' duration). This portfolio is unfortunate because it is maximally exposed to shape risk, i.e., the risk of a non-parallel term structure change. This is because the very long bond is extremely sensitive to changes in the long rate, but the short bond is insensitive. In addition, such a \bar-bell" portfolio may satisfy the two constraints today, but as soon as the short-maturity bond matures (say, within one year), assets and liabilities are suddenly terribly duration mismatched. A related problem is that the portfolio's duration changes rapidly when iterest rates change { that is, the derivative of the duration is large, so if rates change, the asset/liability sides rapidly become duration-mismatched.

5.5 Convexity

Duration matching protects against small, parallel shifts in interest rates, so that the asset and liability sides stay approximately balanced. Mathematically, we are providing a rst-order t between the present-value functions of the assets and liabilities. But more is needed to avoid \bar-bell"-type solutions. This is where convexity enters. 32

as

Convexity (a similar concept used historically is \dispersion", see 9]) is de ned

Qi =

X
2

and is the second derivative, d2 Pi =dri2 , of present value with respect to yields. It turns out that the \bar-bell" portfolio has a very high convexity (known as \positive convexity", according to Fabozzi because it's something good!) By adding to the duration-matched portfolio the requirement that its convexity should be as small as possible, we obtain a portfolio with cash ows that are more centralized in time. But the convexity should not be too small: Intuitively because if the rst bond cash ows arrive after the rst liability we have to borrow, but there's a better, mathematical reason having to do with larger shifts in yields: By requiring that the asset convexity is no less than the liability convexity, we assure that if interest rates change, not only will the asset and liability values change roughly equally, but to the extend they di er, the assets will be worth more! (Think of two curves that rst-order match at some point, but assets have a higher curvature than liabilities. At both sides of the point, assets have the higher value). In other words, we require the net convexity (of assets minus liabilities) to be positive: This is what we mean by \positive net convexity" (and this is why it is \good"!): If rates do change, the asset side comes out ahead of the liability side. The above discussion then leads to the model which minimizes asset convexity, subject to 3 conditions: (1) present value match, (2) duration match, and (3) positive net convexity.
(Immunization Model with Convexity) Minimize P Q x
x U
2

t T

t(t + 1)Fit (1 + ri )

(t+2)

Immun-2:
i

P Subject to Pi U Pi xi = PL dol Didol xi = DL Pi U Q x QL i U i i

2 2 2 2

i U i

xi

i2U

Be aware that this whole discussion centers on parallel shifts and resulting smooth changes in present values. Many derivative xed-income instruments such as options, mortgage-backed securities, etc., have non-smooth response functions, which may make a local immunization unreliable. The current state-of-the art response to this problem is the use of interest-rate scenarios leading to stochastic models, or the use of formal term-structure models which, through pricing and cash- ow models, can be used for hedging and immunization, for instance in a multi-stage, stochastic program. 33

References: See Zenios, Ch. 1, 3.1{3.2 6] for a complete statement of the immunization and factor models. Chapter 31 in 9], by P.E. Christensen and F.J. Fabozzi contains a good discussion of immunization, with examples. Tuckman, Part 3, 31] contains a more rigorous exposition, including Key Duration Matching (Part 2 outlines the modern approach of using a term structure model to rate changes). A more introductory approach is given in Bodie, Kane and Markus 1998], Ch. 16, 4] and similar texts. Exercises: (1) Find the Macaulay-duration of a T -year, zero-coupon bond. (2) Show that the Macaulay-duration of a perpetuity (i.e., a bond paying only coupon payments, forever) with coupon c, using a constant discount rate r, is (1 + r)=r.
The simple immunization model Immun-1 can be extented to directly address other types of term structure changes than parallel shifts. The underlying concept is to model each type of change as a factor, or a vector that speci es how interest rates change. For instance, the rst factor may be

5.6 Factor Immunization

a1 t = (1 1 1 1 ::: 1)T 2 <30

which, when added in some (positive or negative) multiple to the term structure corresponds to a parallel up- or down-shift a second factor might be

a2 t = (;15 ;24 ;13 ;12 ::: ;1 0 1 ::: +13 +14)T 2 <30

which indicates a twist in the structure, and a third might be

a3 t = (0 0:5 1 ::: 4 4:5 4 3:5 ::: 0:5 0)T 2 <30

which is a kind of curvature change. Assume that a number j 2 J such factors are given by aj t. We can now build a model which explicitly hedges against these types of term structure changes. Let yt be the term structure. Then a bond price is given by

Pi =
Di erentiate w.r.t. each yt :

X
2

t T

Fi t (1 + yt ) t :
;

(33)

dFi = ; X F (1 + y ) it t dyt t T
2

(t+1)

drt :

(34)

34

We can formalize the meaning of the aj t 's by noting that some multiple of them, Fj is added to yt , hence the change in the term structure yt is given by:

dyt =

X
2

j J

aj t dFj

(35)

which, substituted into (34), yields an expression for

j t T
2

X dP fi j = dFi = ; t aj t Fi t (1 + yt)

(t+1)

(36)

This number, fi j , is the sensitivity of bond i to changes in the term structure, as given by factor j , or the bond's factor loading. If we can build an asset portfolio which has the same sensitivity to all factors as the liability side, we are immunized against changes given by the factors. The model is now straightforward:
(Factor Immunization Model) Maximize ; P Ddol r x
x U
2

Immun-3:
i

Subject to

P P x i Pi U f i x i U ij i
2 2 2

i U i

xi

= PL = fL j 0

j2J i2U

where we again maximize (an approximation to) the portfolio's yield. Note that there is no constraint matching dollar duration (as in Immun-1). This can be achieved by including the rst factor, a1 t , shown above. This model also comprises key-duration matching models which seek to immunize against changes in individual rates, y 1 y 2 :::, by having vectors aj t with just a single non-zero each (at 1 2 :::). This model is very exible because the decision maker can include any factor of interest. But how many factors are \enough"? Dahl 5] has comments on this. In a study on US data, as few as 3 factors explained 99.9% of the term structure variation (these were level, steepness, and curvature factors), and in Denmark, 4 factors explained 99.6% of the variation: steepness, curvature, a short-maturity factor (up to 10 years), and a very short-term (4 years) twist-factor. The most signi cant factors, based on historical data, were found by statistical methods (principal component analysis).

35

A Algebraic Formulation of Stochastic Network Model

Below is given the complete, algebraic formulation of the stochastic model described in Section 3.3.1, and the notation and variables of the model are de ned.

Two-Stage Stochastic Model for Funding SPDA Liabilities: 1X Maximize S U (W s=E ) (37) Xs s Subject to y0i + us ; v1 =C for all s 2 S , (38) 1 i X s s s ((1 ; )zpi ; ypi) ; us+1 + (1 + rp)us p p i s s s ;(1 + rp + )vp + vp+1 = Ls for all s 2 S (39) p p = 1 ::: Y ; 1 X s + (1 + rs )us (1 ; )zY i Y Y i s ;(1 + rY + )vY = Ls + W s for all s 2 S (40) Y s = y0i x1i for all s 2 S i 2 U (41) s s xs ; z s + y s x(p+1)i = mpi pi pi pi for all s 2 S i 2 U (42) p = 1 ::: Y ; 1 s = ms xs zY i for all s 2 S i 2 U : (43) Yi Yi
2S 2U 2U 2U

All variables are non-negative.

Portfolio Construction and Rebalancing Variables:

y0i : Contents of the initial portfolio (in cash value) of instrument i ( rst-stage
s ypi : s zpi :

variables). Purchase of instrument i at time p under scenario s. Sale of instrument i at time p under scenario s.

Auxiliary Variables:
xs : Holdings (in cash value) of instrument i during period p, under scenario s. pi us : Amount of cash invested at the short rate in period p. p
36

s vp : Amount of cash borrowed in period p. W s: Final wealth under scenario s.

Model Data:
C : The initial cash invested, consisting of the SPDA premium, P , and equity, E. s rp: The short rate during period p under scenario s. ms : The change in value of a $1 holdings of instrument i in period p under pi scenario s. Ls : The liability due at time p under scenario s. p
: Transaction cost (as a fraction of the transaction). : The spread between borrowing and reinvestment rates.

37

References
1] E. Adamidou, Y. Ben-Dov, L. Pendergast, and V. Pica. The optimal portfolio system { targeting horizon total returns under varying interest-rate scenarios. In S.A. Zenios, editor, Financial Optimization. Cambridge University Press, Cambridge, U.K., 1992. 2] J.R. Birge. Decomposition and partitioning methods for multistage stochastic linear programs. Operations Research, 33(5):989{1007, 1985. 3] J.R. Birge. Stochastic programming: Computation and applications. INFORMS Journal on Computing, 9(2):111{133, 1997. 4] Bodie, Kane, and Marcus. Investments. Irwin, 1989. 5] H. Dahl. A exible approach to interest-rate risk management. In S.A. Zenios, editor, Financial Optimization. Cambridge University Press, Cambridge, U.K., 1992. 6] H. Dahl, A. Meeraus, and S.A. Zenios. Some nancial optimization models: I. risk management. In S.A. Zenios, editor, Financial Optimization, pages 3{36. Cambridge University Press, 1993. 7] G.B. Dantzig and G. Infanger. Large-scale stochastic linear programs: Importance sampling and Benders decomposition. Technical report, Operations Research Department, Stanford University, Stanford, CA, 1991. Report SOL 91{4. 8] R. Dembo. Scenario immunization. In S.A. Zenios, editor, Financial Optimization. Cambridge University Press, Cambridge, U.K., 1992. 9] F.J. Fabozzi and I.M. Pollack, editors. The Handbook of Fixed Income Securities. Dow-Jones, Irwin, 1987. 10] H.I. Gassmann. MSLiP: A computer code for the multistage stochastic linear programming problem. Math. Prog., 47:407{423, 1990. 11] J.C. Hull. Options, Futures, and other Derivative Securities. Prentice Hall, Englewood Cli s, New Jersey 07632, 1993. Second Edition. 12] G. Infanger. Planning under Uncertainty: Solving large-Scale Stochastic Linear Programs. Scienti c Press, Danvers, MA, 1994. 13] Jr. Ingersoll, J.E. Theory of Financial Decision Making. Studies in Financial Economics. Rowman & Little eld, 1987. 38

14] F. Macaulay. Some theoretical problems suggested by the movements of interest rates, bond yields and stock prices in the United States since 1856. Technical report, National Bureau of Economic Research, New York, 1938. 15] L.C. MacLean, W.T. Ziemba, and G.Blazenko. Growth versus security in dynamic investment analysis. Management Science, 38(11):1562{1585, Nov. 1992. 16] J. Mulvey. Incorporating transaction costs in models for asset allocation. In S.A. Zenios, editor, Financial Optimization. Cambridge University Press, Cambridge, U.K., 1992. 17] J.M. Mulvey. Network models for cash ow analysis. Journal of Cash Flow Management, 4(1), 1984. 18] J.M. Mulvey. Nonlinear network models in nance. Advances in Mathematical Programming and Financial Planning, 1987. JAI Press. 19] J.M. Mulvey and A. Ruszczynski. A diagonal quadratic approximation method for large scale linear programs. Operations Research Letters, 12:205{ 215, 1992. 20] J.M. Mulvey and H. Vladimirou. Stochastic network optimization models for investment planning. Annals of Operations Research, 20:187{217, 1989. 21] J.M. Mulvey and H. Vladimirou. Stochastic network programming for nancial planning problems. Management Science, 38(11):1642{1664, 1992. 22] J.M. Mulvey and S.A. Zenios. Diversifying xed-income portfolios: modeling dynamic e ects. Financial Analysts Journal, pages 30{38, January/February 1994. 23] S.S. Nielsen and E.I Ronn. A two-factor model for the valuation of the Tbond futures contract's embedded options. In F. Fabozzi, editor, Advances in Fixed-Income Valuation Modeling and Risk Management, pages 135{152, 1995. 24] S.S. Nielsen and S.A. Zenios. Solving multistage stochastic network programs on massively parallel computers. Mathematical Programming, 73:227{250, 1996. 25] S.S. Nielsen and S.A. Zenios. A stochastic programming model for funding Single Premium Deferred Annuities. Mathematical Programming, Series B, 75:177{200, 1996. 39

26] S.S. Nielsen and S.A. Zenios. Scalable parallel Benders decomposition for stochastic linear programming. Parallel Computing, 23:1069{1088, 1997. 27] D.R. Cari no, D.H. Myers, and W.T. Ziemba. Concepts, technical issues, and uses of the russell-yasuda kasai nancial planning model. Technical report, Frank Russell Company, Jan. 1997. 28] R.T. Rockafellar and R.J.-B. Wets. Scenarios and policy aggregation in optimization under uncertainty. Math. of Operations Research, 16(1):119{147, February 1991. 29] A. Ruszczynski. Parallel decomposition of multistage stochastic programming problems. Working paper WP{88{094, IIASA, October. 1988. 30] A. Ruszczynski. Regularized decomposition and augmented Lagrangian decomposition for angular linear programming problems. Aspiration Based Decision Support Systems, 1989. A. Levandowski, A. P. Wierzbicki (Eds). Springer Verlag. 31] Tuckman. Fixed Income Securities. John Wiley & Sons, 1995. 32] R.J-B. Wets. Stochastic programs with xed resources: The equivalent deterministic problem. SIAM Review, 16:309{339, 1974.

40