LIFE INSURANCE MATHEMATICS 2002

Ragnar Norberg London School of Economics Abstract Since the pioneering days of Black, Merton and Scholes nancial mathematics has developed rapidly into a ourishing area of science. Its impacts on insurance are great by any calculation: applications are virtually countless and even the basic paradigms are being rethought. This talk focuses on life insurance and shows how the mathematics of nance and of insurance dovetail into a consistent, model-based approach to measurement and management of combined insurance risk and nance risk.

Introduction

Finance was always an essential part of insurance. Trivially, one might say, because any business has to attend to its money aairs. However, for at least two reasons insurance is not just any business. In the rst place, its products are not physical goods or services; they are nancial contracts with provisions related to uncertain future events. Therefore, pricing of insurance products is not just an accounting exercise involving the four basic arithmetical operations; it is a matter of risk assessment based on stochastic models and methods. In the second place, insurance policies are more or less long term contracts (in life insurance up to several tens of years) under which the customers pays in advance for benets to come later, hence the term premium (= rst) for the price. Therefore, the insurance industry is a major accumulator of capital in todays society, and the insurance companies are major institutional investors. It follows that the nancial operations of an insurance company may be as decisive of its revenues as its insurance operations and that the nancial risk (or asset risk) may be as severe as, or even more severe than, the insurance risk (or liability risk). Insurance risk, which is due to the random nature of the cash-ow of premiums less insurance claims, is diversiable in the sense that gains and losses on individual policies will average out in a suciently large portfolio of independent risks: the law of large numbers is at work. Financial risk, which is due to the uncertain yields on the nancial investments, is not diversiable in the same simple sense. Any investment portfolio is aected by booms and recessions of the market in large, interest rate variations, and movements in prices of individual stocks. The composition of an investment portfolio may be more or less risky (e.g. stocks are more risky than government bonds), but the volume of the portfolio is of course of no relevance to its riskiness; losses and gains on investments do not subject themselves to the law of large numbers.

On this background one may ask why insurance mathematics traditionally centers on measurement and control of the insurance risk. The answer may partly be found in institutional circumstances: The insurance industry used to be heavily regulated, solvency being the primary concern of the regulatory authority. Possible adverse developments of economic factors (ination, weak returns on investment, low interest rates, etc.) would be safeguarded against by placing premiums on the safe side. The surpluses, which would typically accumulate under this regime, were redistributed as bonuses (dividends) to the policyholders only in arrears, after interest and other nancial parameters had been observed. Furthermore, the insurance industry used to be separated from other forms of business and protected from competition within itself, and severe restrictions were placed on its investment operations. In these circumstances nancial matters appeared to be something the traditional actuary did not need to worry about. Another reason why insurance mathematics used to be void of nancial considerations was, of course, the absence of a well developed theory for description and control of nancial risk. All this has changed. National and institutional borders have been downsized or eliminated and regulations have been liberalized: Mergers between insurance companies and banks are now commonplace, new insurance products are being created and put on the market virtually every day, by insurance companies and other nancial institutions as well, and without prior licencing by the supervisory authority. The insurance companies of today nd themselves placed on a ercely competitive market. Many new products are directly linked to economic indices, like unit-linked life insurance and catastrophe derivatives. By so-called securitization also insurance risk can be put on the market and thus open new possibilities of inviting investors from outside to participate in risk that previously had to be shared solely between the participants in the insurance insurance schemes. These developments in practical insurance coincide with the advent of modern nancial mathematics, which has equipped the actuaries with a well developed theory within which nancial risk and insurance risk can be analyzed, quantied and controlled. A new order of the day is thus set for the actuarial profession. The purpose of this chapter is to give a glimpse into some basic ideas and results in modern nancial mathematics and to indicate by examples how they may be applied to actuarial problems involving management of nancial risk.

Classical life insurance mathematics

A. The multi-state life insurance policy. We consider an insurance policy issued at time 0 and terminating at a xed nite time T . There is a nite set of mutually exclusive states of the policy, Z = {1, . . . , J Z }. By convention, 1 is the initial state at time 0. Let Z(t) denote the state of the policy at time t [0, T ]. Taking Z to be a stochastic process with right-continuous paths and at most a nite number of jumps, the same holds also for the associated indicator processes Ij and counting processes Njk 2

Z dened, respectively, by Ij (t) = 1[Z(t) = j] (1 or 0 according as the policy is in Z the state j or not at time t) and Njk (t) = { ; Z( ) = j, Z( ) = k, (0, t]} (the number of transitions from state j to state k (= j) during the time interval (0, t]). The history of the policy up to and including time t is represented by Z the sigma-algebra Ft = {Z( ) ; [0, t]}. The development of the policy is Z given by the ltration (increasing family of sigma-algebras) FZ = {Ft }t[0,T ] . Let B(t) denote the total amount of contractual benets less premiums payable during the time interval [0, t]. We assume that it develops in accordance with the dynamics

dB(t) =
j

Z Ij (t) dBj (t) + j=k

Z bjk (t) dNjk (t) ,

(2.1)

where each Bj is a deterministic payment function specifying payments due during sojourns in state j (a general life annuity), and each bjk is a deterministic function specifying payments due upon transitions from state j to state k (a general life assurance). We assume that each Bj is nite-valued, right-continuous, and decomposes into an absolutely continuous part and a discrete part with at most a nite number of jumps in [0, T ]. Thus, dBj (t) = bj (t)dt + Bj (t) , where Bj (t) = Bj (t) Bj (t), when dierent from 0, is a jump representing a lump sum payable at time t if the policy is then in state j. The functions bj and bjk are assumed to be nite-valued and piecewise continuous. B. The Markov chain description of the policy. The process Z is assumed to be a time-continuous Markov chain on the state space Z. We denote its transition probabilities by pZ (t, u) = P[Z(u) = k | Z(t) = j] , jk t u, and the corresponding intensities of transition by jk (t) = lim pZ (t, t + h)/h , jk
h0

j = k, implying that these exist for all t [0, n). We assume, moreover, that the intensities are piece-wise continuous. The total intensity of transition from state j is j = k; k=j jk . Here and elsewhere a dot in the place of a subscript signies summation over that subscript. We shall need Kolmogorovs forward dierential equations for s < t: dt pZ (s, t) = ij
g; g=j Z We remind of the fact that the compensated counting processes Mjk , j = k, dened by Z Z Z dMjk (t) = dNjk (t) Ij (t)jk (t) dt ,

pZ (s, t) gj (t) dt pZ (s, t) j (t) dt . ig ij

(2.2)

(2.3)

are square integrable, mutually orthogonal, zero mean FZ -martingales. Let Hjk , j, k Z, j = k, be predictable processes satisfying E
j=k (0,t] 2 Hjk ( ) jk ( )d < .

(2.4)

Then the stochastic integral M (t) =

j=k Z Hjk (t) dMjk (t)

is a zero mean square integrable FZ -martingale. Figure 1 outlines the disability model, which is apt to describe a policy on a single life, with payments depending on the state of health of the insured (e.g. a life assurance with waiver of premium during disability).

a active J

i 12  21

invalid

J 13 ^ J d dead

23


Figure 1: Sketch of a Markov chain model for disabilities, recoveries, and death. C. Interest. We assume that the investment portfolio of the insurance company bears interest with intensity r(t) at time t. Thus, for a given realization of r(), the t value at time t of a unit payable at time is e r if t (an accumulation or compounding factor) and e t r if t (a discounting factor). The shorthand r = r(s)ds will be in use throughout. We assume that r is piecewise T continuous and that 0 r is nite. D. The reserve. By statute the insurer must currently maintain a reserve to meet future liabilities in respect of the contract. Since these liabilities are unknown, the reserve can only be an estimate based on the information currently available. The reserve

at time t is dened presicely as the conditional expected present value of the future benets less premiums, given the past history of the policy: V (t) = E
t

By the Markov assumption, V (t) = VZ(t) (t) =

j Z Ij (t) Vj (t) ,

where the Vj are the statewise reserves, Vj (t) = E

t

e e

=
t

r g

We need the backward (so-called Thieles) dierential equations dVj (t) = r(t)Vj (t) dt dBj (t)
k; k=j

where Rjk (t) = bjk (t) + Vk (t) Vj (t) (2.9)

is the (for evident reasons so-called) sum at risk associated with a possible transition from state j to state k at time t. The dierential equation is easily derived by dierentiating the dening expression (2.6) and using the Kolmogorov equation (2.2). E. The principle of equivalence. The contractual payments (or rather the premiums for given benets) are constrained by the principle of equivalence, which lays down that E
0

that is, V1 (0) = B1 (0) . (2.11)

The rationale of this principle, which is the basic paradigm of classical life insurance, is that it will establish balance on the average in a large portfolio of independent policies. 5

dB( ) Z(t) = j

pjg (t, ) dBg ( ) +

dB( )

dB( ) Ft .

(2.5)

h;h=g

bgh ( )gh ( ) d .(2.6) (2.7)

Rjk (t) jk (t) dt ,

(2.8)

= 0,

(2.10)

The reserve dened in the previous paragraph is motivated the same way: upon providing currently a reserve equal to the conditional expected value of the future net liability on each individual reserve, the company will meet its liabilities on the average. F. A martingale proof of Thieles dierential equation. We take here the opportunity to demonstrate a martingale technique that will be omnipresent throughout the text. Thieles dierential equation could be proved by elementary calculations as explained above, but the following argument gives also an insight into the dynamics of the total cash-ow. Dene the martingale M (t) = E
0

e e e

=
0 t

dB( ) + e dB( ) + e

E
t

=
0

t 0

r j

Z Ij (t)Vj (t) .

Now apply Its formula to (2.12): o dM (t) = e +e

r

dB(t) + e
r j

Z Ij (t) dVj (t)

+e

The last term on the right takes care of the jumps of the Markov process: upon a jump from state j to state k the last term in (2.12) changes immediately from the discounted value of the reserve in state j just before the jump to the value of the reserve in state k at the time of the jump. Since the state-wise reserves are deterministic functions with nite variation, they have at most a countable number of discontinuities at xed times. The probability that the Markov process jumps at any such time is 0. Therefore, we need not worry Z about possible common points of discontinuity of the Vj (t) and the Ij (t). For the same reason we can also disregard the left limit in Vj (t) in the last term. Z Z We proceed by inserting (2.1) for dB(t) and the expression dNjk (t) = dMjk (t) Z Ij (t) jk (t) dt obtained from (2.3), and gather dM (t) = e
t 0

+ e

t 0

Z Ij (t) dBj (t) r(t) Vj (t) dt + dVj (t) + Z Rjk (t) dMjk (t) ,

j=k

Since the last term on the right of (2.13) is the increment of a martingale, the rst term of the right is the dierence between the increments of two martingales 6

t 0

t 0

t 0

(r(t) dt)
j
t 0

Z Ij (t) Vj (t) r j=k Z dNjk (t) (Vk (t) Vj (t)) .

Z dB( ) Ft
t 0

Z dB( ) Ft

(2.12)

k; k=j

jk dt Rjk (t)

(2.13)

and is thus itself the increment of a martingale. This martingale has nite variation and, as will be explained below, is also continuous, and must therefore be constant. For this to be true for all realizations of the indicator functions Z Ij , we must have dBj (t) rj Vj (t) dt + dVj (t) +
k; k=j

jk dt Rjk (t) = 0 ,

(2.14)

which is Thieles dierential equation. Furhtermore we obtain that dM (t) = e

t 0

r j=k

Z Rjk (t) dMjk (t) ,

which displays the dynamics of the martingale M . Finally, we explain why the (2.14) is the increment at t of a continuous function. The dt terms are continuous increments, of course. Outside jump times of the Bj both the Bj themselves and the Vj are continuous. At any time t where there is a jump in some Bj the reserve Vj jumps by the same amount in the opposite direction since Vj (t) = Bj (t) + Vj (t). Thus, Bj + Vj is indeed continuous.

Insurance risk

A. Dierential equations for moments of present values. We want to determine higher order moments of future net liabilities. By the Markov property, we need only the state-wise conditional moments =E
t

q = 1, 2, . . . (q) The functions Vj are determined by the dierential equations d (q) (q) (q1) V (t) = (qr(t) + j (t))Vj (t) qbj (t)Vj (t) dt j q q (qp) jk (t) (t))p Vk (t) , p jk p=0
k; k=j

valid on (0, T )\D and subject to the conditions

q

Vj (t) =
p=0

(q)

q (qp) (Bj (t) Bj (t))p Vj (t) , p

t D. A rigorous proof is given in [25]. The computation goes as follows. First solve the dierential equations in the upper interval (tm1 , n), where the side conditions (3.1) are just Vj (n) = (Bj (n) Bj (n))q 7
(q)

(q) Vj (t)

q
t

dB( )

Z(t) = j ,

(3.1)

(3.2)

since Vj (n) = q0 (the Kronecker delta). Then, if m > 1, solve the dierential equations in the interval (tm2 , tm1 ) subject to (3.1) with t = tm1 , and proceed in this manner downwards. (q) Letting mj (t) denote the q-th central moment corresponding to the noncentral Vj (t), we have mj (t) mj (t)
(q) (1) (q)

(q)

= Vj (t) ,
q

(1)

(3.3) q (p) (1) V (t) Vj (t) p j

qp

=
p=0

(1)qp

(3.4)

B. Numerical examples. Referring to disability model, consider a male insured at age 30 for a period of 30 years. We assume that the intensities of transition at time t, when the insured is 30 + t years old, are ad (t) = id (t) = 0.0005 + 0.000075858 100.038(30+t) , ai (t) = 0.0004 + 0.0000034674 100.06(30+t) , ia (t) = 0.005 , and that the interest rate is constant, r = 0.05. (q) The central moments mj (0) dened in (3.3) (3.4) have been computed for the states a and i (state d is uninteresting) and are shown in Table 1 for (1) a term insurance with sum 1 (= b02 = b12 ); (2) an annuity payable in active state with level intensity 1 (= b0 ); (3) an annuity payable in disabled state with level intensity 1 (= b1 ); (4) for a combined policy providing a term insurance with sum 1 (= b02 = b12 ) and a disability annuity with level intensity 0.5 (= b1 ) against level net premium 0.0125 (= b0 ) payable in active state.

Table 1: Central moments mj (0), q = 1, 2, 3, of the present value of the four payments streams (1) (4) listed above. ma (0) (1) (2) (3) (4) 0.063 14.748 0.245 0.000
(1)

(q)

mi (0) 0.063 0.790 13.573 7.158

(1)

ma (0) 0.027 4.816 1.477 0.397

(2)

mi (0) 0.027 6.510 8.675 2.234

(2)

ma (0) 0.012 39.540 12.061 1.597

(3)

mi (0) 0.012 61.399 69.004 9.379

(3)

C. Solvency margins. Let W be the present value of all future net liabilities in respect of an insurance portfolio. Denote the q-th central moment of W by m(q) . The so-called normal

power approximation of the upper -fractile of the distribution of W , which we denote by w1 , is based on the rst three moments and is w1 m(1) + c1 m(2) + c2 1 m(3) 1 , 6 m(2)

where c1 is the upper -fractile of the standard normal distribution. Adopting the so-called break-up criterion in solvency control, w1 can be taken as a minimum requirement on the technical reserve at the time of consideration. It decomposes into the premium reserve, m(1) , and what can be termed the uctuation reserve, w1 m(1) . A possible measure of the riskiness of the portfolio is the ratio R = w1 m(1) /P , where P is some suitable measure of the size of the portfolio. By way of illustration, consider a portfolio of N independent policies, all identical to the combined policy (4) in Table 1. Taking as P the total premium income per year, the value of R at the time of issue is 48.61 for N = 10, 12.00 for N = 100, 3.46 for N = 1000, 1.06 for N = 10000, and 0.332 for N = 100000.

Stochastic interest and nancial risk.

A. A Markov chain interest model. The economy (or rather the part of the economy that governs the interest) is modeled as a homogeneous time-continuous Markov chain Y on a nite state space Y = {1, . . . , J Y }, with intensities of transition ef , e, f J Y , e = f . Y Y The associated indicator and counting processes are denoted by Ie and Nef , Y respectively, and the ltration generated by Y is denoted by FY = {Ft }t[0,T ] . We assume that the force of interest takes a xed value re when the economy is in state e, that is, r(t) = rY (t) =
e Y Ie (t)re .

(4.1)

Figure 2 outlines a simple Markov chain interest rate model with three states {1, 2, 3} and state-wise rates of interest r1 = 0.02 (low), r2 = 0.05 (medium), and r3 = 0.08 (high). Direct transition can only be made to a neighbouring state, and the total intensity of transition out of any state is 0.5, that is, the interest rate changes every two years on the average. By symmetry, the stationary (long run average) interest rate is 0.05. B. The full Markov model. We assume that the processes Y and Z are deined on the same probability space and that they are stochastically independent. Then (Y, Z) is a Markov chain on Y Z with intensities ef (t) , e = f, j = k , jk (t) , e = f, j = k , ej,f k (t) = 0, e = f, j = k . 9

r1 = 0.02

12 = 0.5  21 = 0.25

r2 = 0.05

23 = 0.25  32 = 0.5

r3 = 0.08

Figure 2: Sketch of a simple Markov chain interest model. For the purpose of assessing the contractual liability we are interested in aspects of its conditional distribution, given the available information at time t. We focus here on determining the conditional moments. By the Markov assumption, the functions in quest are the state-wise conditional moments =E
t

The dierential equations extend straightforwardly to d (q) (q) (q1) V (t) = (qre + j (t) + e )Vej (t) qbj (t)Vej (t) dt ej q q (qp) (q) (bjk (t))p Vek (t) ef Vf j (t) . jk (t) p p=0
k;k=j f ;f =e

Denote by mej (t) the q-th central moment corresponding to Vej (t). C. Numerical results for a combined insurance policy. Consider the combined life insurance and disability pension policy and the Markov chain interest model in Figure 2 modied such that the innitesimal matrix is 1 1 0 = 0.5 1 0.5 . (4.3) 0 1 1 The scalar can be interpreted as the expected number of interest changes per time unit. Table 2 displays the rst three central moments of the present value at time 0. The level premium rate (= b1 ) is the net premium rate in state (2,a) (i.e. the rate that establishes expected balance between discounted premiums and benets when the insured is active and the interest is at medium level 0.05 at time 0). The rst three rows in the body of the table form a benchmark; = 0 means no interest uctuation, and we therefore obtain the results for three cases of xed interest. It is seen that the second and third order moments of the 10

(q)

(q) Vej (t)

q
t

dB( )

Y (t) = e, Z(t) = j .

(4.2)
(q)

Table 2: Central moments mej (0) of orders q = 1, 2, 3 of the present value of future benets less premiums for the combined policy in interest state e and policy state j at time 0, for some dierent values of the rate of interest changes, . Second column gives the net premium of a policy starting from interest state 2 (medium) and policy state 1 (active). e, j : q 0.072 10.206 1.156 6.024 6.845 36.093 0.030 9.166 0.838 5.622 4.652 21.814 0.001 7.601 0.467 3.104 2.070 10.546 0.000 0.404 1.643 0.000 0.397 1.597 7.207 2.330 9.491 7.158 2.234 9.379 0.000 7.158 0.397 2.234 1.597 9.379 0.000 7.254 0.438 3.151 1.895 7.923 0.000 7.212 0.416 2.767 1.740 8.808 0.000 7.165 0.399 2.302 1.613 9.322 0.000 7.158 0.397 2.234 1.597 9.379 0.030 5.318 0.159 0.938 0.465 2.947 0.018 5.843 0.236 1.793 0.806 3.110 0.001 6.854 0.370 2.467 1.462 7.357 0.000 7.123 0.394 2.274 1.585 9.155 0.000 7.158 0.397 2.234 1.597 9.379 1, 1 1, 2 2, 1 2, 2 3, 1 3, 2

(q)

1 .0125 2 3

1 .05 .0127 2 3 1 .0126 2 3 1 .0125 2 3 1 .0125 2 3

.5

11

present value are strongly dependent on the (xed) force of interest and, in fact, their absolute values decrease when the force of interest increases (as could be expected since increasing interest means decreasing discount factors and, hence, decreasing present values of future amounts). It is seen that, as increases, the dierences across the three pairs of columns get smaller and in the end they vanish completely. The obvious interpretation is that the initial interest level is of little importance if the interest changes rapidly. The overall impression from the two central columns corresponding to medium interest is that, as increases from 0, the variance of the present value will rst increase to a maximum and then decrease again and stabilize. This observation supports the following piece of intuition: the introduction of moderate interest uctuation adds uncertainty to the nal result of the contract, but if the interest changes suciently rapidly, it will behave like xed interest at the mean level. Presumably, the values of the net premium in the second column reect the same eect.

Getting rid of nancial risk

A. Non-diversiable risk. The principle of equivalence rests on the implicit assumption that the experience basis, that is the transition intensities, interest, and administration costs throughout the contract period, are known at the time of inception of the contract. In reality, however, the experience basis may undergo signicant and unforeseeable changes within the time horizon of the contract, thus exposing the insurer to a risk that is non-diversiable. In the present paper, which focuses on the interplay between insurance and nance, we will concentrate on the interest rate and assume that all other elements of the experience basis are known and xed throughtout the term of the contract. The risk stemming from the uncertain development of the interest rate can, under certain ideal market conditions, be eliminated by letting the contractual payments depend on the returns on the companys investments. Products of this type, known as unit-linked insurances, have been gaining increasing market shares ever since they emerged some few decades ago, and today they are also theoretically well understood, see Aase and Persson (1994), Mller (1998), and references therein. Unlike the unit-linked concept, a standard life insurance policy species contractual payments in nominal amounts, binding to both parties throughout the entire term of the contract. Thus, an adverse development of the interest rates can not be countered by raising premiums or reducing benets and also not by cancelling the contract (the right of withdrawal remains one-sidedly with the insured). The only way the insurer can prevent the non-diversiable nancial risk is to charge premiums to the safe side. In practice this is done by calculating premiums on a conservative so-called technical basis or rst order basis, which represents a provisional worst-case scenario for the future development

12

of the experience basis. In our simplied set-up, with interest as the only uncertain element in the experience basis, this means that premiums and reserves are calculated under the assumption that the interest rate is r , typically lower than expected. Denote the corresponding rst order state-wise reserve by V j . Premiums are set in accordance with the principle of equivalence, E
0

or, equivalently, V1 (0) = B1 (0) . (5.2) B. Denition of the surplus. With premiums based on prudent rst order assumptions, the portfolio will create a systematic technical surplus if everything goes well. We dene the surplus at time t as S(t) =
0

e
t 0

= e

e
0

which is past net income (premiums less benets), compounded with the factual second order interest, minus expected discounted future liabilities valuated on the conservative rst order basis. This denition complies with practical accountancy regulations in insurance since S(t) is precisely the dierence between the current cash balance and the rst order reserve that by statute has to be provided to meet future liabilities. Notice that S(0) = 0, a consequence of (5.1), T T and S(T ) = 0 e dB( ), as it ought to be. Dierentiating (5.4) gives dS(t) = e
r

r(t) dt
0

= r(t) dt S(t) + r(t) dt VZ(t) (t) dB(t) dVZ(t) (t) . Upon substituting dB(t) from (2.1) and VZ(t) (t) from (2.5), using the general It formula to write dVZ(t) (t) = j Z Ij (t) dVj (t) + j=k Z {Vk (t) Vj (t)} dNjk (t)

(there are almost surely no common jumps of the deterministic state-wise reserves and the counting processes), and picking dVj (t) from (2.8), we nd dS(t) = (t) dt S(t) + dC(t) + dM (t) , where dC(t) =
j Z Ij (t) cj (t) dt ,

13

t 0

dB( ) = 0 ,

(5.1)

d(B)( ) VZ(t) (t) t

0

(5.3) (5.4)

dB( ) VZ(t) (t) ,

dB( ) dB(t) dVZ(t) (t)

(5.5)

with cj (t) = (r(t) r ) Vj (t) and dM (t) =

j=k Z with the Mjk dened in (2.3). The process M is a zero mean H-martingale in the conditional model, given Gn , that is, E[M (t) | Hs Gn ] = M (s) Z Rjk (t) dMjk (t) ,

(5.6)

for s t, and M (0) = 0. Then it is also a zero mean F-martingale in the full model since E[M (t) | Fs ] = E [ E[M (t) | Hs Gn ] | Fs ] = E[M (s) | Fs ] = M (s). The term dM (t) in (5.5) is the purely accidental part of the surplus increment. The two rst terms on the right of (5.5) are the systematic parts, which make the surplus drift to something with expected value dierent from 0. The rst term is the earned interest on the surplus itself, and what remains is quite naturally the policy-holders contribution to the technical surplus. To put it another way, let us switch the rst term on the right of (5.5) over t to the left and multiply the equation with e 0 to form a complete dierential on the left hand side. Integrating from 0 to t and using the fact that S(0) = C(0) = M (0) = 0, we arrive at e

S(t) =
0

dC( ) +
0

showing that the discounted surplus at time t is the discounted total contributions plus a martingale representing noise. We can rewrite (5.7) as S(t) =
0

dC( ) +
0

displaying the surplus at time t as the compounded total of contributions and accidental martingale increments. (Beware that the last term on the right of (5.8) is not a martingale although its expected value is 0 for all t.) C. Redistribution of surplus Bonus. The technical surplus belongs to the insured and has to paid back as bonus. There are many possible schemes. The simplest is cash bonus: The rate at which bonus will be paid at some xed future time u, provided the insured is then alive, is W = (r(u) r )V (u) . Adopting the Markov chain interest model, we can make model-based prediction of this quantity. At time t < u, given r(t) = re , W is predicted by its conditional expected value We (t) = E[W | r(t) = re ] . 14

dM ( ) ,

t 0

dM ( ) ,

(5.7)

(5.8)

It is easy to show that the functions We (t) are the solution to the dierential equations d We (t) = ef (We (t) Wf (t)) , dt
;f =e

subject to the conditions

We (u) = (re r )Vu ,

e = 1, . . . , J Y . By terminal bonus the surpluses are accumulated and paid back as a lump sum at the term of the contract T , provided the insured is then alive. The amount paid is W =
0

= W (t)
0

where W (t) = e W (t) =

t
T t

The random variables W (t) and W (t), which are unknown at time t, are predicted by We (t) We (t) Writing W (t) W (t) = er(t) dt W (t + dt) , = W (t) (r(t) r ) V (t) dt + W (t + dt) , = E[W (t) | r(t) = re ] , = E[W (t) | r(t) = re ] .

we easily show by a backward argument that the functions We (t) and We (t) are the solution to the dierential equations d W (t) dt e d W (t) dt e = re We (t) +
f ;f =e

= We (t)(re r )V (t) +
f ;f =e

subject to the conditions We (T ) = 1 , We (T ) = 0 , e = 1, . . . , J .

Y

(r( ) r )V ( ) d
t

(r( ) r )V ( ) d + W (t) ,

,
T

(r( ) r )V ( ) d .

ef (We (t) Wf (t)) , ef (We (t) Wf (t)) ,

15

Reintroducing nancial risk, and eliminating it again

A. Guaranteed interest. Recall the basic rules of the with prot insurance contract: On the one hand, any surplus is to be redistributed to the insured. On the other hand, benets and premiums set out in the contract cannot be altered to the insureds disadvantage. This means that negative surplus, should it occur, cannot result in negative bonus. Thus, the with prot policy comes with an interest rate guarantee to the eect that bonus is to be paid as if factual interest were no less than rst order interest, roughly speaking. For instance, cash bonus is to be paid at rate (r(t) r )+ V ( t) per survivor at time t, hence the insurer has to cover (r r(t))+ V (t) . (6.1)

Similarly, terminal bonus (typical for e.g. a pure endowment benet) is to be paid as a lump sum e
0

(r( ) r )V ( ) d
+

per survivor at time n, hence the insurer has to cover e

0

(r r( ))V ( ) d
+

(6.2)

(We write a+ = max(a, 0) = a 0.) An interest guarantee of this kind represents a liability on the part of insurer. It cannot be oered for free, of course, but has to be compensated by a premium. This can certainly be done without violating the rules of game for the participating policy, which lay down that premiums and benets be set out in the contract at time 0. Thus, for simplicity, suppose a single premium is to be collected at time 0 for the guarantee. The question is, how much should it be? Being brought up with the principle of equivalence, we might think that the expected discounted value of the liability is an agreeable candidate for the premium. However, the rationale of the principle of equivalence, which was to make premiums and benets balance on the average in an innitely large portfolio, does not apply to nancial risk. Interest rate variations cannot be eliminated by increasing the size of the portfolio; all policy-holders are faring together in one and the same boat on their once-in-a-lifetime voyage through the troubled waters of their chapter of economic history. This risk cannot be averaged out in the same way as the risk associated with the lengths of the individual lives. 16

None the less, in lack of anything better, let us nd the expected discounted value of the interest guarantee, and just anticipate here that this actually would be the correct premium in an extended model specifying a so-called complete nancial market. Those who are familiar with basic arbitrage theory know what this means. Those who are not should at this stage just imagine that, in addition to the bank account with the interest rate r(t), there are some other investment opportunities, and that any future nancial claim can be duplicated perfectly by investing a certain amount at time 0 and thereafter just selling and buying available assets without any further infusion of capital. The initial amount required to perform this duplicating investment strategy is, quite naturally, the price of the claim. It turns out that this price is precisely the expected discounted value of the claim, only under a dierent probability measure than the one we have specied in our physical model. With these reassuring phrases, let us proceed to nd the expected discounted value of the interest guarantee. B. Pricing guaranteed interest. Consider cash bonus with gurantee given by (6.1): Given that r(0) = re (say), the price of the total claims under the guarantee, averaged over an innitely large portfolio, is E
0

A natural starting point for creating some useful dierential equations by the backward construction is the price of future claims under the guarantee in state e at time t, We (t) = E
t

e = 1, . . . , J Y , 0 t T . The price in (6.3) is precisely We (0). Conditioning on what happens in the time interval (t, t + dt] and neglecting terms of order o(dt) that will disappear in the end anyway, we nd We (t) = (1e dt) (r re )+ V (t) t px dt + ere dt We (t + dt) +
f ; f =e

From here we easily arrive at the dierential equations d We (t) = (r re )+ V (t) t px +re We (t) dt ef dt (Wf (t)We (t)) , (6.5)
f ; f =e

which are to be solved subject to the conditions We (T ) = 0 . (6.6)

(r r( ))+ V ( ) px d r(0) = re .

(6.3)

(r r( ))+ V ( ) px d r(t) = re ,

(6.4)

ef dt Wf (t) .

17

Next, consider terminal bonus at time n given by (6.2): Given r(0) = re , the price of the claim under the guarantee, averaged over an innitely large portfolio, is E e
r 0

e e

=E
0

The price of the claim at time t should be the conditional expected discounted value of the claim, given what we know at the time: E e
r 0

e
n

=E

U (t) +
t

where U (t) =
0

The quantity in (6.8) is more involved than the one in (6.4) since it depends eectively on the past history of interest rate through U (t). We can, therefore, not hope to end up with the same simple type of problem as above and in all other situations encountered so far, where we essentially had to determine the conditional expected value of some function depending only on the future course of the interest rate. Which was easy since, by the Markov property, we could look at state-wise conditional expected values We (t), e = 1, . . . , J Y , say. These are deterministic functions of the time t only and can be determined by solving ordinary dierential equations. Let us proceed and see what happens. Due to the Markov property (conditional independence between past and future, given the present) the expression in (6.8) is a function of t, r(t) and U (t). Dropping the uninteresting factor T px , consider its value for given U (t) = u and r(t) = re , We (t, u) = E u+
t

Use the backward construction: We (t, u) = (1 e dt)E u + (r re )V (t) dt + ere dt

t+dt

18

T t

(r r( ))V ( ) d
+
t

n 0

(r r( ))V ( ) d
+

T px

r(0) = re
T px .

(r r( ))V ( ) d
+

r(0) = re

(6.7)

T px

r( ); 0 t

(r r( ))V ( ) d
+

r( ); 0 t

T px

(6.8)
t
t

(r r( ))V ( ) d .

(r r( ))V ( ) d
+

r(t) = re .

t+dt

(r r( ))V ( ) d
+

r(t + dt) = re

+
f ; f =e

ef dt Wf (t, u) = ef dt Wf (t, u) .
f ; f =e

(1 e dt)ere dt We (t + dt, ere dt u + (r re )Vt dt) + Insert here er

e

dt

= 1 re dt + o(dt),

We (t + dt, ere dt u + (r re )V (t) dt) = We (t, u) + We (t, u) dt + We (t, u)(u re + (r re )V (t)) dt + o(dt) , t u and proceed in the usual manner to arrive at the partial dierential equations We (t, u)+(ure +(r re )V (t)) We (t, u)re We (t, u)+ t u These are to be solved subject to the conditions We (T , u) = u+ , e = 1, . . . , J Y . Since the functions we are interested in involved both t and U (t), we are lead to state-wise functions in two arguments and, therefore, quite naturally end up with partial dierential equations for those. ef (Wf (t, u)We (t, u)) = 0 .
f ; f =j

A Markov chain nancial market

A. Motivation. The theory of diusion processes, with its wealth of powerful theorems and model variations, is an indispensable toolkit in modern nancial mathematics. The seminal papers of Black and Scholes [6] and Merton [21] were crafted with Brownian motion, and so were most of the almost countless papers on arbitrage pricing theory and its bifurcations that followed over the past quarter of a century. A main course of current research, initiated by the martingale approach to arbitrage pricing ([15] and [16]), aims at generalization and unication. Today the core of the matter is well understood in a general semimartingale setting, see e.g. [9]. Another course of research investigates special models, in particular various Levy motion alternatives to the Brownian driving process, see e.g. [10] and [27]. Pure jump processes have been widely used in nance, ranging from plain Poisson processes introduced in [22] to quite general marked point processes, see e.g. [4]. And, as a pedagogical exercise, the market driven by a binomial process has been intensively studied since it was launched in [8]. We will here present a model where the nancial market is driven by a continuous time homogeneous Markov chain. The idea was launched in [26] and 19

reappeared in [11], the context being limited to modelling of the spot rate of interest. This will allow us to synthesize insurance and nance within the mathematical model framework already familiar to us. Some additional notation and results are presented in Appendix H. B. The Markov chain market We are going to extend the interest model in 4. Thus let {Yt }t0 be a continuous time Markov chain with nite state space Y = {1, . . . , J Y }. Recall that Y Y the associated indicator and counting processes are denoted by Ie and Nef . Y Y The F = {Ft }t0 , is taken to satisfy the usual conditions of right-continuity (Ft = u>t Fu ) and completeness (F0 contains all subsets of P-nullsets), and F0 is assumed to be the trivial (, ). This means, essentially, that Y is rightY Y continuous (hence the same goes for the Ie and the Nef ) and that Y0 deterministic. We assume that Y is time homogeneous so that the transition probabilities pY (s, t) = P[Yt = f | Ys = e] ef depend only on the length of the transition period, ts. Henceforth we therefore write pY (s, t) = pY (t s). This implies that the transition intensities ef ef ef = lim
t

pY (t) ef , 0 t

(7.1)

e = f , exist and are constant. To avoid repetitious reminders of the type e, f Y, we reserve the indices e and f for states in Y throughout. We will frequently refer to Ye = {f ; ef > 0} , the set of states that are directly accessible from state e, and denote the number of such states by Y Je = |Ye | . Put ee = e =
f ;f Ye

ef

(minus the total intensity of transition out of state e). We assume that all states intercommunicate so that pY (t) > 0 for all e, f (and t > 0). This implies that ef Y Je > 0 for all e (no absorbing states). The matrix of transition probabilities, PY (t) = (pef (t)) , and the innitesimal matrix, = (ef ) , are related by (7.1), which in matrix form reads = lim t 0 1 (PY (t) I), and t by the backward and forward Kolmogorov dierential equations, d Y P (t) = PY (t) = PY (t) . dt 20 (7.2)

Under the side condition PY (0) = I, (7.2) integrates to PY (t) = exp(t) . In the representation (H.2),
n

(7.3)

PY (t) = De=1,...,J Y (ee t ) 1 =

e=1

e e t j j ,

(7.4)

the rst (say) eigenvalue is 1 = 0, and corresponding eigenvectors are 1 = 1 and 1 = (p1 , . . . , pn ) = limt (pe1 (t), . . . , peJ Y (t)), the stationary distribution of Y . The remaining eigenvalues, 2 , . . . , n , are all strictly negative so that, by (7.4), the transition probabilities converge exponentially to the stationary distribution as t increases. We now turn to the subject matter of our study and, referring to introductory texts like [5] and [32], take basic notions and results from arbitrage pricing theory as prerequisites. C. The continuous time Markov chain market. We consider a nancial market driven by the Markov chain described above. Y Thus, Yt represents the state of the economy at time t, Ft represents the information available about the economic history by time t, and FY represents the ow of such information over time. In the market there are m + 1 basic assets, which can be traded freely and frictionlessly (short sales are allowed, and there are no transaction costs). A special role is played by asset No. 0, which is a locally risk-free bank account with state-dependent interest rate r(t) = rYt =
e

Ie (t)re ,

where the state-wise interest rates re , e = 1, . . . , J Y , are constants. Thus, its price process is
t t

S0 (t) = exp
0

r(s) ds

= exp
e

re
0

Ie (s) ds

with dynamics dS0 (t) = S0 (t) r(t) dt = S0 (t)

e

re Ie (t) dt .

(Setting S0 (0) = 1 a just a matter of convention.) The remaining m assets, henceforth referred to as stocks, are risky, with price processes of the form
t

Si (t) = exp

ie

Ie (s) ds +

f Ye

ief Nef (t) ,

(7.5)

21

i = 1, . . . , m, where the ie and ief are constants and, for each i, at least one of the ief is non-null. Thus, in addition to yielding state-dependent returns of the same form as the bank account, stock No. i makes a price jump of relative size ief = exp (ief ) 1 upon any transition of the economy from state j to state k. By the general Its formula, its dynamics is given by dSi (t) = Si (t) ie Ie (t) dt +
e e f Ye

Taking the bank account as numeraire, we introduce the discounted stock prices Si (t) = Si (t)/S0 (t), i = 0, . . . , m. (The discounted price of the bank account is Bt 1, which is certainly a martingale under any measure). The discounted stock prices are with dynamics Si (t) = exp
t

ief dNef (t) .

(7.6)

(ie re )

Ie (s) ds +

e f Ye

ief Nef (t) ,

(7.7)

i = 1, . . . , m. We stress that the theory we are going to develop does not aim at explaining how the prices of the basic assets emerge from supply and demand, business cycles, investment climate, or whatever; they are exogenously given basic entities. (And God said let there be light, and there was light, and he said let there also be these prices.) The purpose of the theory is to derive principles for consistent pricing of nancial contracts, derivatives, or claims in a given market. D. Portfolios. A dynamic portfolio or investment strategy is an m + 1-dimensional stochastic process (t) = ((t), (t)) , where (t) represents the number of units of the bank account held at time t, and the i-th entry in (t) = (1 (t), . . . , m (t)) represents the number of units of stock No. i held at time t. As it will turn out, the bank account and the stocks will appear to play dierent parts in the show, the latter being the more visible. It is, therefore, convenient to costume

dSi (t) = Si (t)

(ie re )Ie (t) dt +

e e f Ye

ief dNef (t) ,

(7.8)

22

the two types of assets and their corresponding portfolio entries accordingly. To save notation, however, it is useful also to work with double notation (t) = (0 (t), . . . , m (t)) , with 0 (t) = (t), i (t) = i (t), i = 1, . . . , m, and work with St = (S1 (t), . . . , Sm (t)) . and S = (S0 (t), . . . , Sm (t)) . t The portfolio is adapted to FY (the investor cannot see into the future), and the shares of stocks, , must also be FY -predictable (the investor cannot, e.g. upon a sudden crash of the stock market, escape losses by selling stocks at prices quoted just before and hurry the money over to the locally risk-free bank account.) The value of the portfolio at time t is
m

V (t) = (t)S0 (t) +

i=1

i (t)Si (t) = (t)S0 (t) + (t)S(t) = (t)S (t)

Henceforth we will mainly work with discounted prices and values and, in accordance with (7.7), equip their symbols with a tilde. The discounted value of the portfolio at time t is V (t) = (t) + (t) S(t) = (t) S (t) . (7.9)

The strategy is self-nancing (SF) if dV (t) = (t) dS (t) or, equivalently,

m

dV (t) = (t) dS (t) =

i=1

i (t) dSi (t) .

(7.10)

We explain the last step: Put D(t) = S0 (t)1 , a continuous process. The dy namics of the discounted prices S (t) = D(t)S (t) is then dS (t) = dD(t)S (t)+ (t) = D(t)V (t), we have D(t) dS (t). Thus, for V dV (t) = dD(t) V (t) + D(t) dV (t) = dD(t) (t) S (t) + D(t) (t) dS (t) = (t) (dD(t)S (t) + D(t) dS (t)) = (t) dS (t) ,

hence the property of being self-nancing is preserved under discounting. The SF property says that, after the initial investment of V0 , no further investment inow or dividend outow is allowed. In integral form: V (t) = V0 +
t 0

(s) dS(s) = V0 +

t 0

(s) dS(s) .

(7.11)

Obviously, a constant portfolio is SF; its discounted value process is V (t) = S (t), hence (7.10) is satised. More generally, for a continuous port folio we would have dVt () = d (t) S (t) + (t) dS (t), and the self-nancing 23

 condition would be equivalent to the a budget constraint d (t) S (t) = 0, which says that any purchase of assets must be nanced by a sale of some other assets. We urge to say that we shall typically be dealing with portfolios that are not continuous and, in fact, not even right-continuous so that d(t) is meaningless (integrals with respect to the process are not well dened). E. Absence of arbitrage. An SF portfolio is called an arbitrage if, for some t > 0, V0 < 0 and V (t) 0 a.s. P , or, equivalently, V0 < 0 and V (t) 0 a.s. P .

A basic requirement on a well-functioning market is the absence of arbitrage. The assumption of no arbitrage, which appears very modest, has surprisingly far-reaching consequences as we shall see. A martingale measure is any probability measure P that is equivalent to P (t) are martingales (F, P). The and such that the discounted asset prices S fundamental theorem of arbitrage pricing says: If there exists a martingale measure, then there is no arbitrage. This result follows from easy calculations starting from (7.11): Forming expectation E under P and using the martingale under P, we nd property of S E[V (t)] = V0 + E[
t 0

(s) dS(s)] = V0

(the stochastic integral is a martingale). It is seen that arbitrage is impossible. We return now to our special Markov chain driven market. Let = (ef ) be an innitesimal matrix that is equivalent to in the sense that ef = 0 if and only if ef = 0. By Girsanovs theorem, there exists a measure P, equivalent to Consequently, P, under which Y is a Markov chain with innitesimal matrix . Y the processes Mef , j = 1, . . . , J Y , f Ye , dened by Y dMef (t) = dNef (t) Ie (t)ef dt , (7.12)

and Mef (0) = 0, are zero mean, mutually orthogonal martingales w.r.t. (FY , P). Rewrite (7.8) as i = 1, . . . , m. The discounted stock prices are martingales w.r.t. (FY , P) if and only if the drift terms on the right vanish, that is, ie re +
f Ye

i dSi (t) = St

ie re +

f Ye

ief ef Ie (t) dt +

e f Ye

Y , ief dMef (t)(7.13)

ief ef = 0 ,

(7.14)

24

j = 1, . . . , J Y , i = 1, . . . , m. From general theory it is known that the existence of such an equivalent martingale measure P implies absence of arbitrage. The relation (7.14) can be cast in matrix form as re 1 e = e e , j = 1, . . . , J Y , where 1 is m 1 and e = (ie )i=1,...,m ,
e e = (ief )i=1,...,m ,

(7.15)

f Y

e = ef

f Ye

The existence of an equivalent martingale measure is equivalent to the existence of a solution e to (7.15) with all entries strictly positive. Thus, the market is arbitrage-free if (and we can show only if) for each j, re 1 e is in the interior of the convex cone of the columns of e . Assume henceforth that the market is arbitrage-free so that (7.13) reduces to dSi (t) = Si (t)
e f Ye

Y ief dMef (t) ,

(7.16)

where the Mef dened by (7.12) are martingales w.r.t. (FY , P) for some measure that is equivalent to P. P Inserting (7.16) into (7.10), we nd that is SF if and only if
m

dV (t) =
e f Ye i=1

Y i (t)Si (t)ief dMef (t) ,

(7.17)

implying that V is a martingale w.r.t. (FY , P) and, in particular, V (t) = E[V (t) | Ft ] . (7.18)

Here E denotes expectation under P. (Note that the tilde, which in the rst place was introduced to distinguish discounted values from the nominal ones, is also attached to the equivalent martingale measure and certain related entities. This usage is motivated by the fact that the martingale measure arises from the discounted basic price processes, roughly speaking.) F. Attainability. Y A T -claim is a contractual payment due at time T . Formally, it is an FT measurable random variable H with nite expected value. The claim is attainable if it can be perfectly duplicated by some SF portfolio , that is, VT = H . (7.19)

If an attainable claim should be traded in the market, then its price must at any time be equal to the value of the duplicating portfolio in order to avoid

25

arbitrage. Thus, denoting the price process by (t) and, recalling (7.18) and (7.19), we have (t) = V (t) = E[H | Ft ] , or

(7.20)

(t) = E e

T t

H Ft .

(7.21)

By (7.20) and (7.17), the dynamics of the discounted price process of an attainable claim is
m

d (t) =
e f Ye i=1

Y i (t)Si (t)ief dMef (t) .

(7.22)

G. Completeness. Any T -claim H as dened above can be represented as H = E[H] +

0 T e f Ye

Y ef (t)dMef (t) ,

(7.23)

where the ef (t) are FY -predictable and integrable processes. Conversely, any random variable of the form (7.23) is, of course, a T -claim. By virtue of (7.19), and (7.17), attainability of H means that H = V0 + = V0 +
T 0 T 0 e f Ye i

dV (t) Y i (t)Si (t)ief dMef (t) . (7.24)

Comparing (7.23) and (7.24), we see that H is attainable i there exist predictable processes 1 (t), . . . , m (t) such that
m

i (t)Si (t)ief = ef (t) ,

i=1 Y for all j and f Ye . This means that the Je -vector

e (t) = (ef (t))f Ye is in R(e ). The market is complete if every T -claim is attainable, that is, if every nj Y vector is in R(e ). This is the case if and only if rank(e ) = Je , which can be Y fullled for each e only if m maxe Je .

26

Arbitrage-pricing of derivatives in a complete market

A. Dierential equations for the arbitrage-free price. Assume that the market is arbitrage-free and complete so that prices of T -claims are uniquely given by (7.20) or (7.21). Let us for the time being consider a T -claim of the form H = h(Y (T ), S (T )) . (8.1)

Examples are a European call option on stock No. dened by H = (S (T ) K)+ , a caplet dened by H = (r(T ) g)+ = (rYT g)+ , and a zero coupon T -bond dened by H = 1. For any claim of the form (8.1) the relevant state variables involved in the conditional expectation (7.21) are t, Y (t), S (t), hence (t) is of the form
JY

(t) =
e=1

Ie (t)fe (t, S (t)) ,

(8.2)

where the

fe (t, s) = E e

T t

H Y (t) = e, S (t) = s

(8.3)

are the state-wise price functions. The discounted price (7.20) is a martingale w.r.t. (FY , P). Assume that the functions fe (t, s) are continuously diferentiable. Using It on
JY

(t) = e we nd d (t) = e
t 0

t 0

r e=1

Ie (t)fe (t, S (t)) ,

(8.4)

r e

Ie (t) re fe (t, S (t)) +

r e f Ye

fe (t, S (t)) + fe (t, S (t))S (t) t s

ef ))

dt

+e = e +
t 0

t 0

(ff (t, S (t)(1 + Ie (t) (re fe (t, S (t)) +

fe (t, S (t))) dNef (t)

r e

fe (t, S (t)) + fe (t, S (t))S (t) t s

{ff (t, St (1 +
f Ye
t 0

ef ))

fe (t, S (t))}ef dt
ef ))

+e

r e f Ye

(ff (t, S (t)(1 +

Y fe (t, S (t))) dMef (t) .

(8.5)

27

By the martingale property, the drift term must vanish, and we arrive at the non-stochastic partial dierential equations re fe (t, s) + +
f Ye

fe (t, s) + fe (t, s)s e t s (ff (t, s(1 + ef )) fe (t, s)) ef = 0 ,

(8.6)

e = 1, . . . , J Y , which are to be solved subject to the side conditions fe (T, s) = h(e, s) , e = 1, . . . , J Y . In matrix form, with R = Dj=1,...,J Y (re ) , A = Dj=1,...,J Y ( e ) , (8.7)

and other symbols (hopefully) self-explaining, the dierential equations and the side conditions are Rf (t, s) + f (t, s) + sA f (t, s) + f (t, s(1 + )) = 0 , t s f (T, s) = h(s) . (8.8) (8.9)

B. Identifying the strategy. Once we have determined the solution fe (t, s), e = 1, . . . , J Y , the price process is known and given by (8.2). The duplicating SF strategy can be obtained as follows. Setting the drift term to 0 in (8.5), we nd the dynamics of the discounted price; d (t) = e
t 0

r e f Ye

(ff (t, S (t)(1 +

ef ))

Y fe (t, S (t))) dMef (t) .(8.10)

Identifying the coecients in (8.10) with those in (7.22), we obtain, for each state j, the equations
m

i (t)Si (t)ief = ff (t, S (t)(1 +

i=1

ef ))

fe (t, S (t)) ,

(8.11)

f Ye . The solution e (t) = (i,e (t))i=1,...,m (say) certainly exists since rank(e ) m, and it is unique i rank(e ) = m. Furthermore, it is a function of t and S(t) and is thus predictable. This simplistic argument works on the open Y intervals between the jumps of the process Y , where dMef (t) = Ie (t)ef dt. For the dynamics (8.10) and (7.22) to be the same also at jump times, the coecients must clearly be left-continuous. We conclude that
JY

(t) =
e=1

Ie (t)(t) ,

28

which is predictable. Finally, is determined upon combining (7.9), (7.20), and (8.4):
JY m

(t) = e

t 0

r e=1

Ie (t)fe (t, S (t)) Ie (t)

i=1

i,e (t)Si (t)

C. The Asian option. As an example of a path-dependent claim, let us consider an Asian option, which essentially is a T -claim of the form H = The price process is (t) =
e

T 0

S ( ) d K
+

, where K 0.

= E e

T t

S ( ) d K
t

Y Ft

Ie (t)fe t, S (t),
0

S ( ) d

where fe (t, s, u) =

The discounted price process is (t) = e

t 0

E e

T t

r t

S ( ) + u K

Y (t) = j, S (t) = s . (8.12)

t

n r j=1

Ie (t) fe t, S (t),
0

(s)

We obtain partial dierential equations in three variables. The special case K = 0 is simpler, with only two state variables. D. Interest rate derivatives. A particularly simple, but still important, class of claims are those of the form H = h(YT ). Interest rate derivatives of the form H = h(rT ) are included since r(T ) = rYT . For such claims the only relevant state variables are t and Y (t), so that the function in (8.3) depends only on t and e. The equation (8.6) reduces to d fe (t) = re fe (t) dt (ff (t) fe (t))ef ,
f Ye

(8.13)

and the side condition is (put h(e) = he ) fe (T ) = he . In matrix form, d f (t) = (R )f (t) , dt 29 (8.14)

subject to f (T ) = h . The solution is f (t) = exp{( R)(T t)}h . (8.15)

It depends on t and T only through T t. In particular, the zero coupon bond with maturity T corresponds to h = 1. We will henceforth refer to it as the T -bond in short and denote its price process by p(t, T ) and its state-wise price functions by p(t, T ) = (pe (t, T ))e=1,...,J Y ; p(t, T ) = exp{( R)(T t)}1 . (8.16)

For a call option on a U -bond, exercised at time T (< U ) with price K, h has entries he = (pe (T, U ) K)+ . In (8.15) (8.16) it may be useful to employ the representation shown in (H.2),
exp{( R)(T t)} = Dj=1,...,J Y (ej (T t) ) 1 ,

(8.17)

say.

Numerical procedures

A. Simulation. The homogeneous Markov process {Y (t)}t[0,T ] is simulated as follows: Let K be the number of transitions between states in [0, T ], and let T1 , . . . , TK be the successive times of transition. The sequence {(Tn , Y (Tn ))}n=0,...,K is generated recursively, starting from the initial state Y (0) at time T0 = 0, as follows. Having arrived at Tn and Y (Tn ), generate the next waiting time Tn+1 Tn as an exponential variate with parameter Y (n) (e.g. ln(Un )/Y (n) , where Un has a uniform distribution over [0, 1]), and let the new state Y (Tn+1 ) be k with probability Y (n)k /Y (n) . Continue in this manner K+1 times until TK < T TK+1 . B. Numerical solution of dierential equations. Alternatively, the dierential equations must be solved numerically. For interest rate derivatives, which involve only ordinary rst order dierential equations, a Runge Kutta will do. For stock derivatives, which involve partial rst order dierential equations, one must employ a suitable nite dierence method, see e.g. [35].

10

Risk minimization in incomplete markets

A. Incompleteness. The notion of incompleteness pertains to situations where a contingent claim 30

cannot be duplicated by an SF portfolio and, consequently, does not receive a unique price from the no arbitrage postulate alone. In Paragraph ??F we were dealing implicitly with incompleteness arising from a scarcity of traded assets, that is, the discounted basic price processes are incapable of spanning the space of all martingales w.r.t. (FY , P) and, in particular, reproducing the value (7.23) of every nancial derivative (function of the basic asset prices). Incompleteness also arises when the contingent claim is not a purely nancial derivative, that is, its value depends also on circumstances external to the nancial market. We have in mind insurance claims that are caused by events like death or re and whose claim amounts are e.g. ination adjusted or linked to the value of some investment portfolio. In the latter case we need to work in an extended model specifying a basic probability space with a ltration F = {Ft }t0 containing FY and satisfying the usual conditions. Typically it will be the natural ltration of Y and some other process that generates the insurance events. The denitions and conditions laid down in Paragraphs ??C-E are modied accordingly, so that adaptedness of and predictability of are taken to be w.r.t. (F, P) (keeping the symbol P for the basic probability measure), a T -claim H is FT measurable, etc. B. Risk minimization. Throughout the remainder of the paper we will mainly be working with discounted prices and values without any other mention than the notational tilde. The reason is that the theory of risk minimization rests on certain martingale representation results that apply to discounted prices under a martingale measure. We will be content to give just a sketchy review of some main concepts and results from the seminal paper of Fllmer and Sondermann [12]. Let H be a T -claim that is not attainable. This means that an admissible portfolio satisfying V (T ) = H cannot be SF. The cost, C (t), of the portfolio by time t is dened as that part of the value that has not been gained from trading: C (t) = V (t)
0 t

( )dS( ) .

The risk at time t is dened as the mean squared outstanding cost, R(t) = E (C (T ) C (t))2 Ft . By denition, the risk of an admissible portfolio is R (t) = E (H V (t)
t T

(10.1)

( )dS( ))2 Ft ,

which is a measure of how well the current value of the portfolio plus future trading gains approximates the claim. The theory of risk minimization takes 31

this entity as its object function and proves the existence of an optimal admissible portfolio that minimizes the risk (10.1) for all t [0, T ]. The proof is constructive and provides a recipe for how to actually determine the optimal portfolio. One sets out by dening the intrinsic value of H at time t as V H (t) = E H | Ft . Thus, the intrinsic value process is the martingale that represents the natural current forecast of the claim under the chosen martingale measure. By the Galchouk-Kunita-Watanabe representation, it decomposes uniquely as V H (t) = E[H] +
0 t

H (t)dS(t) + LH (t) ,

where LH is a martingale w.r.t. (F, P) which is orthogonal to S. The portfolio H dened by this decomposition minimizes the risk process among all admissible strategies. The minimum risk is RH (t) = E
t T

d LH ( ) Ft .

C. Unit-linked insurance. As the name suggests, a life insurance product is said to be unit-linked if the benet is a certain predetermined number of units of an asset (or portfolio) into which the premiums are currently invested. If the contract stipulates a minimum value of the benet, disconnected from the asset price, then one speaks of unit-linked insurance with guarantee. A risk minimization approach to pricing and hedging of unit-linked insurance claims was rst taken by Mller [23], who worked with the Black-Scholes-Merton nancial market. We will here sketch how the analysis goes in our Markov chain market, which conforms well with the life history process in that they both are intensity-driven. Let Tx be the remaining life time of an x years old who purchases an insurance at time 0, say. The conditional probability of survival to age x + u, given survival to age x + t (0 t < u), is
ut px+t

= P[Tx > u | Tx > t] = e

u t

x+s ds

(10.2)

where y is the mortality intensity at age y. We have d ut px+t = ut px+t x+t dt . Introduce the indicator of survival to age x + t, I(t) = 1[Tx > t] , and the indicator of death before time t, N (t) = 1[Tx t] = 1 I(t) . 32 (10.3)

The process N (t) is a (very simple) counting process with intensity I(t) x+t , that is, M given by dM (t) = dN (t) I(t) x+t dt (10.4)

is a martingale w.r.t. (F, P). Assume that the life time Tx is independent of the economy Y . We will work with the martingale measure P obtained by replacing and leaving the rest of the intensity matrix of Y with the martingalizing the model unaltered. Consider a unit-linked pure endowment benet payable at a xed time T , contingent on survival of the insured, with sum insured equal to one unit of stock No. , but guaranteed no less than a xed amount g. This benet is a contingent T -claim, H = (S (T ) g) I(T ) . The single premium payable as a lump sum at time 0 is to be determined. Let us assume that the nancial market is complete so that every purely nancial derivative has a unique price process. Then the intrinsic value of H at time t is V H (t) = (t) I(t) T t px+t , where (t) is the discounted price process of the derivative S (T ) g. Using It and inserting (10.4), we nd dV H (t) = d (t) I(t) = d (t) I(t)
T t px+t T t px+t

+ (t) I(t)

T t px+t x+t

dt + (0 t T t px+t ) dN (t)

t T t px+t dM (t) .

It is seen that the optimal trading strategy is that of the price process of the sum insured multiplied with the conditional probability that the sum will be paid out, and that dLH (t) = T t px+t t dM (t) . Consequently, RH (t) =
t T 2 T s px+s E T

(s)2 Ft

st px+t

x+s ds x+s ds . (10.5)

T t px+t t

E (s)2 Ft

T s px+s

11

Trading with bonds: How much can be hedged?

A. A nite zero coupon bond market. Suppose an agent faces a contingent T -claim and is allowed to invest only in the bank account and a nite number m of zero coupon bonds with maturities Ti , i = 1, . . . , m, all post time T . For instance, regulatory constraints may be imposed on the investment strategies of an insurance company. The question is, to what extent can the claim be hedged by self-nanced trading in these available assets? 33

 An allowed SF portfolio has discounted value process V (t) of the form

m

dV (t) =
i=1

i (t)
e f Ye

Y (f (t, Ti )e (t, Ti ))dMef (t) = p p

e

d(Me (t)) Fe (t)(t) ,

Y where is predictable, MY (t) = (Mef (t))f Ye is the ne -dimensional row vector e Y comprising the non-null entries in the j-th row of MY (t) = (Mef (t)), and Fe (t) = Ye Ft where
i=1,...,m Ft = (e (t, Ti ))e=1,...,J Y = ( (t, T1 ), , p(t, Tm )) , p p

(11.1)

i=1,...,m Y and Ye is the Je J Y matrix which maps Ft to (f (t, Ti ) pe (t, Ti ))f Ye . p

If e.g. Y n = {1, . . . , p}, then Y J = (Ipp , 0p(np1) , 1p1 ). The sub-market consisting of the bank account and the m zero coupon bonds is complete in respect of T -claims i the discounted bond prices span the space of all martingales w.r.t. (FY , P) over the time interval [0, T ]. This is the case Y Y i, for each e, rank(Fe (t)) = Je . Now, since Ye obviously has full rank Je , the rank of Fe (t) is determined by the rank of Ft in (11.1). We will argue that, typically, Ft has full rank. Thus, suppose c = (c1 , . . . , cm ) is such that Ft c = 0 J Recalling (8.16), this is the same as
m
Y

ci exp{( R)Ti }1 = 0 ,
i=1

or, by (8.17) and since has full rank,

m

Dj=1,...,J Y (
i=1

ci ee Ti )1 1 = 0 .

(11.2)

Since 1 has full rank, the entries of the vector 1 1 cannot be all null. Typically all entries are non-null, and we assume this is the case. Then (11.2) is equivalent to
m c i e e Ti = 0 , i=1

j = 1, . . . , J Y .

(11.3)

Using the fact that the generalized Vandermonde matrix has full rank, we know that (11.3) has a non-null solution c if and only if the number of distinct eigenvalues e is less than m. 34

In the case where rank(Fe (t)) < nj for some j we would like to determine the Y Galchouk-Kunita-Watanabe decomposition for a given FT -claim. The intrinsic value process has dynamics d Ht =
e f Ye

Y ef (t)dMef (t) =
e

d(Me (t)) e (t) .

(11.4)

We seek a decomposition of the form dV (t) =

i

i (t) d(t, Ti ) + p
e f Ye

Y ef (t) dMef (t) Y ef (t)dMef (t)

e f Ye

=
e jYe i

Y i (t) (f (t, Ti ) pe (t, Ti )) dMef (t) + p d(Me (t)) Fe (t) e (t) +

e e

d(Me (t)) e (t) ,

such that the two martingales on the right hand side are orthogonal, that is,
j It e f Ye

(Fe (t) e (t)) e e (t) = 0 ,

where e = D(e ). This means that, for each e, the vector e (t) in (11.4) is to be decomposed into its , e projections onto R(Fe (t)) and its orthocomple ment. From (H.3) and (H.4) we obtain Fe (t) e (t) = Pe (t) e (t) , where hence e (t) = (Fe (t) e Fe (t))1 Fe (t) e e (t) . Furthermore, e (t) = (I Pe (t)) e (t) , and the risk is
T

Pe (t) = Fe (t)(Fe (t) e Fe (t))1 Fe (t) e ,

(11.5)

(11.6)

pY (t)e (s t)
t e f Ye

ef (ef (s))2 ds .

(11.7)

The computation goes as follows: The coecients ef involved in the intrinsic value process (11.4) and the state-wise prices pj (t, Ti ) of the Ti -bonds are obtained by simultaneously solving (8.6) and (8.13), starting from (8.9) and (8.13), respectively, and at each step computing the optimal trading strategy by (11.5) and the from (11.6), and adding the step-wise contribution to the 35

variance (11.7) (the step-length times the current value of the integrand). B. First example: The oorlet. For a simple example, consider a oorlet H = (r rT )+ , where T < mini Ti . The motivation could be that at time T the insurance company will ascribe interest to the insureds account at current interest rate, but not less than a prexed guaranteed rate r . Then H is the amount that must be provided per unit on deposit and per time unit at time T . Computation goes by the scheme described above, with the ef (t) = ff (t) (t) obtained from (8.13) subject to (8.14) with he = (r re )+ . fe C. Second example: The interest guarantee in insurance. A more practically relevant example is an interest rate guarantee on a life insurance policy. Premiums and reserves are calculated on the basis of a prudent so-called rst order assumption, stating that the interest rate will be at some xed (low) level r throughout the term of the insurance contract. Denote the corresponding rst order reserve at time t by V (t). The (portfolio-wide) mean surplus created by the rst order assumption in the time interval [t, t + dt) is (r r(t))+ t p V (t) dt. This surplus is currently credited to the account of the x insured as dividend, and the total amount of dividends is paid out to the insured at the term of the contracts at time T . Negative dividends are not permitted, however, so at time T the insurer must cover H =
0

The intrinsic value of this claim is e

0

Ht

= E
t

s 0

(r r(s))+ s p V (s) ds Ft x
t 0

=
0

s 0

(r r(s))+ s p V (s) ds + e x

where the fe (t) are the state-wise expected values of future guarantees, discounted at time t, e
t

fe (t) = E

s t

(r r(s))+ s p V (s) ds Y (t) = e . x

Working along the lines of Section 8, we determine the fe (t) by solving d fe (t) = (r re )+ t p V (t) + re fe (t) x dt subject to fe (T ) = 0 . 36 (11.8) (ff (t) fe (t))ef ,
f Ye

T s

(r r(s))+ s p V (s) ds . x

r e

Ie (t)fe (t) ,

 The intrinsic value has dynamics (11.4) with ef (t) = ff (t) fe (t). From here we proceed as described in Paragraph A. D. Computing the risk. Constructive dierential equations may be put up for the risk. As a simple example, for an interest rate derivative the state-wise risk is Re (t) =
t T

peg ( t)
g f ;f =g

gf (gf ( )) d .

Dierentiating this equation, we nd d Re (t) = dt ef (ef (t)) +

f ;f =e t g 2 T

d peg ( t) dt

(gf ( )) d ,
f ;f =g

and, using the backward version of (7.2), d peg (s t) = dt we arrive at d Re (t) = dt ef (ef (t))
f ;f =e f ;f =e 2

eh phg (s t) + e peg (s t) ,
h;h=e

ef Rf (t) + e Re (t) .

References
[1] Aase, K.K. and Persson, S.-A. (1994). Pricing of unit-linked life insurance policies. Scand. Actuarial J., 1994, 26-52. [2] Andersen, P.K., Borgan, ., Gill, R.D., Keiding, N. (1993). Statistical Models Based on Counting Processes. Springer-Verlag, New York, Berlin, Heidelberg. [3] Bibby J.M., Mardia, K.V., and Kent J.T. Multivariate Analysis. Academic Press, 1979. [4] Bjrk, T., Kabanov, Y., Runggaldier, W. (1997): Bond market structures in the presence of marked point processes. Mathematical Finance, 7, 211-239. [5] Bjrk, T. (1998): Arbitrage Theory in Continuous Time, Oxford University Press. [6] Black, F., Scholes, M. (1973): The pricing of options and corporate liabilities. J. Polit. Economy, 81, 637-654. [7] Bremaud, P. (1981). Point processes and queues. Springer-Verlag, New York, Heidelberg, Berlin. 37

[8] Cox, J., Ross, S., Rubinstein, M. (1979): Option pricing: A simplied approach. J. of Financial Economics, 7, 229-263. [9] Delbaen, F., Schachermayer, W. (1994): A general version of the fundamental theorem on asset pricing. Mathematische Annalen, 300, 463-520. [10] Eberlein, E., Raible, S. (1999): Term structure models driven by general Lvy processes. Mathematical Finance, 9, 31-53. [11] Elliott, R.J., Kopp, P.E. (1998): Springer-Verlag. Mathematics of nancial markets,

[12] Fllmer, H., Sondermann, D. (1986): Hedging of non-redundant claims. In Contributions to Mathematical Economics in Honor of Gerard Debreu, 205223, eds. Hildebrand, W., Mas-Collel, A., North-Holland. [13] Gantmacher, F.R. (1959): Matrizenrechnung II, VEB Deutscher Verlag der Wissenschaften, Berlin. [14] Gerber, H.U. (1995). Life Insurance Mathematics, 2nd edn. SpringerVerlag. [15] Harrison, J.M., Kreps, D.M. (1979): Martingales and arbitrage in multiperiod securities markets. J. Economic Theory, 20, 1979, 381-408. [16] Harrison, J.M., Pliska, S. (1981): Martingales and stochastic integrals in the theory of continuous trading. J. Stoch. Proc. and Appl., 11, 215-260. [17] Hoem, J.M. (1969): Markov chain models in life insurance. Bltter Deutsch. Gesellschaft Vers.math., 9, 91107. [18] Hoem, J.M. and Aalen, O.O. (1978). Actuarial values of payment streams. Scand. Actuarial J. 1978, 38-47. [19] Jordan, C.W. (1967). Life Contingencies. The Society of Actuaries, Chicago. [20] Karlin, S., Taylor, H. (1975): A rst Course in Stochastic Processes, 2nd. ed., Academic Press. [21] Merton, R.C. (1973): The theory of rational option pricing. Bell Journal of Economics and Management Science, 4, 141-183. [22] Merton, R.C. (1976): Option pricing when underlying stock returns are discontinuous. J. Financial Economics, 3, 125-144. [23] Mller, T. (1998): Risk minimizing hedging strategies for unit-linked life insurance. ASTIN Bull., 28, 17-47. [24] Norberg, R. (1991). Reserves in life and pension insurance. Scand. Actuarial J. 1991, 1-22. 38

[25] Norberg, R. (1995): Dierential equations for moments of present values in life insurance. Insurance: Math. & Econ., 17, 171-180. [26] Norberg, R. (1995): A time-continuous Markov chain interest model with applications to insurance. J. Appl. Stoch. Models and Data Anal., 245-256. [27] Norberg, R. (1998): Vasiek beyond the normal. Working paper No. 152, Laboratory of Actuarial Math., Univ. Copenhagen. [28] Norberg, R. (1999): A theory of bonus in life insurance. Finance and Stochastics., 3, 373-390. [29] Norberg, R. (2001): On bonus and bonus prognoses in life insurance. Scand. Actuarial J. 1991, 1-22. [30] Papatriandafylou, A. and Waters, H.R. (1984). Martingales in life insurance. Scand. Actuarial J. 1984, 210-230. [31] Protter, P. (1990). Stochastic Integration and Dierential Equations. Springer-Verlag, Berlin, New York. [32] Pliska, S.R. (1997): Introduction to Mathematical Finance, Blackwell Publishers. [33] Ramlau-Hansen, H. (1991): Distribution of surplus in life insurance. ASTIN Bull., 21, 57-71. [34] Sverdrup, E. (1969): Noen forsikringsmatematiske emner. Stat. Memo. No. 1, Inst. of Math., Univ. of Oslo. (In Norwegian.) [35] Thomas, J.W. (1995): Numerical Partial Dierential Equations: Finite Dierence Methods, Springer-Verlag. [36] Width, E. (1986): A note on bonus theory. Scand. Actuarial J. 1986, 121126.

Appendices A Calculus

A. Piecewise dierentiable functions. Being concerned with operations in time, commencing at some initial date, we will consider functions dened on the positive real line [0, ). Thus, let us consider a generic function X = {Xt }t0 and think of Xt as the state or value of some process at time t. For the time being we take X to be real-valued. 39

In the present text we will work exclusively in the space of so-called piecewise dierentiable functions. From a mathematical point of view this space is tiny since only elementary calculus is needed to move about in it. From a practical point of view it is huge since it comfortably accommodates any idea, however sophisticated, that an actuary may wish to express and analyse. It is convenient to enter this space from the outside, starting from a wider class of functions. We rst take X to be of nite variation (FV), which means that it is the dierence between two non-decreasing, nite-valued functions. Then the leftlimit Xt = limst Xs and the right-limit Xt+ = limst Xs exist for all t, and they dier on at most a countable set D(X) of discontinuity points of X. We are particularly interested in FV functions X that are right-continuous (RC), that is, Xt = limst Xs for all t. Any probability distribution function is of this type, and any stream of payments accounted as incomes or outgoes, can reasonably be taken to be FV and, as a convention, RC. If X is RC, then Xt = Xt Xt , when dierent from 0, is the jump made by X at time t. For our purposes it suces to let X be of the form
t

Xt = X 0 +
0

x d +
0< t

(X X ) .

(A.1)

The integral, which may be taken to be of Riemann type, adds up the continuous increments/decrements, and the sum, which is understood to range over discontinuity times, adds up increments/decrements by jumps. We assume, furthermore, that X is piecewise dierentiable (PD); A property holds piecewise if it takes place everywhere except, possibly, at a nite number of points in every nite interval. In other words, the set of exceptional points, if not empty, must be of the form {t0 , t1 , . . .}, with t0 < t1 < , and, in case it is innite, limj tj = . Obviously, X is PD if both X and x are piecewise d continuous. At any point t D = D(X) D(x) we have dt Xt = xt , that is, the / function X grows (or decays) continuously at rate xt . As a convenient notational device we shall frequently write (A.1) in dierential form as dXt = xt dt + Xt Xt . (A.2)

A left-continuous PD function may be dened by letting the sum in (A.1) range only over the half-open interval [0, t). Of course, a PD function may be neither right-continuous nor left-continuous, but such cases are of no interest to us. B. Integrals. Let X and Y both be PD and, moreover, let X be RC and given by (A.2). The integral over (s, t] of Y with respect to X is dened as
t t

Y dX =
s s

Y x d +
s< t

Y (X X ) ,

(A.3)

40

provided that the individual terms on the right and also their sum are well dened. Considered as a function of t the integral is itself PD and RC with continuous increments Yt xt dt and jumps Yt (Xt Xt ). One may think of the integral as the weighted sum of the Y -values, with the increments of X as weights, or vice versa. In particular, (A.1) can be written simply as
t

Xt = X s +
s

dX ,

(A.4)

saying that the value of X at time t is its value at time s plus all its increments in (s, t]. By denition,
t t t

Y dX = lim
s r

Y dX =
s s

Y dX Yt (Xt Xt ) =
(s,t)

Y dX ,

a left-continuous function of t. Likewise,

t t t

Y dX = lim
s r

Y dX =
r s

Y dX + Ys (Xs Xs ) =
[s,t]

Y dX ,

a left-continuous function of s. C. The chain rule (Its formula). 1 m Let Xt = (Xt , . . . , Xt ) be an m-variate function with PD and RC components i i i i given by dXt = xt dt + (Xt Xt ). Let f : Rm R have continuous partial derivatives, and form the composed function f (Xt ). On the open intervals where there are neither discontinuities in the xi nor jumps of the X i , the function f (Xt ) develops in accordance with the well-known chain rule for scalar elds along rectiable curves. At the exceptional points f (Xt ) may change (only) due to jumps of the X i , and at any such point t it jumps by f (Xt ) f (Xt ). Thus, we gather the so-called change of variable rule or Its formula, which in our simple function space reads
m

df (Xt ) =
i=1

f (Xt ) xi dt + f (Xt ) f (Xt ) , t xi

(A.5)

or, in integral form,

t m

f (Xt ) = f (Xs ) +
s i=1

f (X ) xi d + xi

{f (X ) f (X )} . (A.6)
s< t

Obviously, f (Xt ) is PD and RC. A frequently used special case is (check the formulas!) d(Xt Yt ) = Xt yt dt + Yt xt dt + Xt Yt Xt Yt = Xt dYt + Yt dXt + (Xt Xt )(Yt Yt ) = Xt dYt + Yt dXt . 41 (A.7)

If X and Y have no common jumps, as is certainly the case if one of them is continuous, then (A.7) reduces to the familiar d(Xt Yt ) = Xt dYt + Yt dXt . The integral form of (A.7) is the so-called rule of integration by parts:
t t

(A.8)

Y dX = Yt Xt Ys Xs
s s

X dY .

(A.9)

Probability and expectation

Taking basic measure theoretic probability as a prerequisite, we represent the relevant part of the world and its uncertainties by a probability space (, F, P). Here is the set of possible outcomes , F is a sigmaalgebra of subsets of representing the events to which we want to assign probabilities, and P : F [0, 1] is a probability measure. A set A F such that P[A] = 0 is called a nullset, and a property that takes place in all of , except possibly on a nullset, is said to hold almost surely (a.s.). If more than one probability measure are in play, we write nullset (P) and a.s. (P) whenever emphasis is needed. Two probability measures P and P, if they are dened on the same F P are said to be equivalent, written P and have the same nullsets. Let G be some sub-sigmaalgebra of F. We denote the restriction of P to G by PG ; PG [A] = P[A], A G. Note that also (, G, PG ) is a probability space. A G-measurable random variable (r.v.) is a function X : R such that X 1 (B) G for all B R, the Borel sets in R. We write X G in short. The expected value of a r.v. X is the probability-weighted average E[X] = X dP = X() dP(), provided this integral is well dened. The conditional expected value of X, given G, is the r.v. E[X|G] G satisfying E{E[X|G] Y } = E[XY ] (B.1)

for each Y G such that the expected value on the right exists. It is unique up to nullsets (P). To motivate (B.1), consider the special case when G = {B1 , B2 , . . .}, the sigma-algebra generated by the F-measurable sets B1 , B2 , . . ., which form a partition of . Being G-measurable, E[X|G] must be of the form k bk 1Bk . Putting this together with Y = 1Bj into the relationship (B.1) we arrive at X dP B E[X|G] = 1 Bj j , P[Bj ] j as it ought to be. In particular, taking X = IA , we nd the conditional probability P[A|B] = P[A B]/P[B]. One easily veries the rule of iterated expectations, which states that, for H G F, E { E[X|G]| H} = E[X|H] . 42 (B.2)

Change of measure

If L is a r.v. such that L 0 a.s. (P) and E[L] = 1, we can dene a probability measure P on F by P[A] =
A

L dP = E[1A L] .

(C.1)

If L > 0 a.s. (P), then P P. The expected value of X w.r.t. P is E[X] = E[XL] (C.2)

if this integral exists; by the denition (C.1), the relation (C.2) is true for indicators, hence for simple functions and, by passing to limits, it holds for measurable functions. Spelling out (C.2) as X dP = XL dP suggests the = L dP or notation dP dP = L. dP The function L is called the Radon-Nikodym derivative of P w.r.t. P. Conditional expectation under P is formed by the rule E[XL|G] E[X|G] = . E[L|G] To see this, observe that, by denition, E{E[X|G] Y } = E[XY ] for all Y G. The expression on the left of (C.5) can be reshaped as E{E[X|G] Y L} = E{E[X|G] E[L|G] Y } . The expression on the right of (C.5) is E[XY L] = E{E[XL|G] Y } . It follows that (C.5) is true for all Y G if and only if E[X|G] E[L|G] = E[XL|G] , which is the same as (C.4). For X G we have EG [X] = E[X] = E[XL] = E {X E[L|G]} = EG {X E[L|G]} , showing that d PG = E[L|G] . dPG 43 (C.7) (C.6) (C.5) (C.4) (C.3)

Stochastic processes: general concepts

To describe the evolution of random phenomena over some time interval [0, T ], we introduce a family F = {Ft }0tT of sub-sigmaalgebras of F, where Ft represents the information available at time t. More precisely, Ft is the set of events whose occurrence or non-occurrence can be ascertained by time t. If no information is ever sacriced, we have Fs Ft for s < t. We then say that F is a ltration, and (, F, F, P) is called a ltered probability space. A stochastic process is a family of r.v.-s, {Xt }0tT . It is said to be adapted to the ltration F if Xt Ft for each t [0, T ], that is, at any time the current state (and also the past history) of the process is fully known if we are currently provided with the information F. An adapted process is said to be predictable if its value at any time is entirely determined by its history in the strict past, loosely speaking. For our purposes it is sucient to think of predictable processes as being either left-continuous or deterministic.

Martingales
E[Xt |Fs ] = Xs

An adapted process X with nite expectation is a martingale if

for s < t. The martingale property depends both on the ltration and on the probability measure, and when these need emphasis, we shall say that X is martingale (F, P). The denition says that, on the average, a martingale is always expected to remain on its current level. One easily veries that, conditional on the present information, a martingale has uncorrelated future increments. Here are some useful general results: Abbreviate Pt = PFt , introduce Lt = and put L = LT . By (C.7) we have Lt = E[L|Ft ] , (E.1) d Pt , dPt

which is a martingale (F, P). Let X be a real-valued random variable such that E|X| < . Then the process M dened by M (t) = E[X|Ft ] is a martingale. This follows by the rule of iterated expectation and the ltration property, Fs Ft for s < t : E[M (t) | Fs ] = E{E[X|Ft ] | Fs } = E[X | Fs ] = M (s) . A martingale M with paths that are (almost surely) continuous and of nite variation in every nite interval is constant as a function of time; M (t) = M (0) 44

for all t. This is seen as follows. Since M has nite variation, it obeys the rules of ordinary calculus and, in particular, M 2 (t) = M 2 (0) + 2
0 t

M (s) dM (s) .
t 0

Since M is continuous, it is also predictable so that the integral is a martingale. It follows that E M 2 (t) = M 2 (0) . Since E[M (t)] = M (0), we conclude that Var[M (t)] = 0 , hence M is constant.

2M (s) dM (s)

Counting processes

As the name suggests, a counting process is a stochastic process N = {Nt }0tT that commences from zero (N0 = 0) and thereafter increases by isolated jumps N N of size 1 only. The natural ltration of N is FN = {Ft }0tT , where Ft = {Ns ; s t} is the history of N by time t. This is the smallest ltration to N which N is adapted. The strict past history of N at time t is denoted by Ft . N An F -predictable process {t }0tT is called a compensator of N if the process M dened by Mt = N t t (F.1)

is a zero mean FN -martingale. If is absolutely continuous, that is, of the form

t

t =
0

s ds ,

then the process is called the intensity of N . We may also dene the intensity informally by t dt = P [dNt = 1 | Ft ] = E [dNt | Ft ] , and we sometimes write the associated martingale (F.1) in dierential form, dMt = dNt t dt . (F.2)

A stochastic integral w.r.t. the martingale M is an FN -adapted process H of the form

t

Ht = H 0 +
0

hs dMs ,

(F.3)

N where H0 is F0 -measurable and h is an FN -predictable process such that H is integrable. The stochastic integral is also a martingale.

45

A fundamental representation result states that every FN martingale is a N stochastic integral w.r.t. M . It follows that every integrable Ft measurable r.v. is of the form (F.3). t (1) t (2) (1) (1) (2) (2) If Ht = H0 + 0 hs dMs and Ht = H0 + 0 hs dMs are stochastic integrals with nite variance, then an easy heuristic calculation shows that Cov[HT , HT |Ft ] = E and, in particular,
T (1) (2) T t

h(1) h(2) s ds | Ft , s s

(F.4)

Var[HT |Ft ] = E
t

h2 s ds | Ft . s

H (1) and H (2) are said to be orthogonal if they have conditionally uncorrelated increments, that is, the covariance in (F.4) is null. This is equivalent to saying that H (1) H (2) is a martingale. The intensity is also called the innitesimal characteristic if the counting process since it entirely determines it probabilistic properties. If t is deterministic, then Nt is a Poisson process. If depends only on Nt , then Nt is a Markov process. The change of variable rule (A.6) becomes particularly simple when the argument is a counting function. In fact, for f : R R and for N a counting process, we have f (Nt ) = f (Ns ) +
s< t

{f (N ) f (N )} {f (N + 1) f (N )}(N N )
s< t t

(F.5) (F.6) (F.7)

= f (Ns ) + = f (Ns ) +
s

{f (N + 1) f (N )} dN .

What these expressions state, is just

j

f (j) = f (0) +
i=1

{f (i) f (i 1)} .

Still they are useful representations when we come to stochastic counting processes. A comprehensive textbook on counting processes in life history analysis is [2].

The Girsanov transform

Girsanovs theorem is celebrated in stochastics, and it is basic in mathematical nance. We formulate and prove the counting process variation: 46

Theorem (Girsanov). Let Nt be a counting process with (F, P)-intensity t , and let t be a given non-negative F-adapted process such that t = 0 if and such that P P and only if t = 0. Then there exists a probability measure P t . The likelihood process (E.1) is N has (F, P)-intensity
t

Lt = exp
0

(ln s ln s ) dNs +
0

(s s ) ds

Proof: We shall give a constructive proof, starting from a guessed L in (C.3). Since L must be strictly positive a.e. (P), a candidate would be L = LT , where
t t

Lt = exp
0

s dNs +
0

s ds

with predictable and adapted. In the rst place, Lt should be a martingale (F, P). By Its formula, dLt = Lt t dt + Lt (et 1) dNt = Lt t + et 1 t dt + Lt et 1 dMt . The representation result (F.3) tells us that to make L a martingale, we must make the drift term vanish, that is, t = 1 e t t , whereby dLt = Lt et 1 dMt , In the second place, we want to determine t such that the process M given by dMt = dNt t dt (G.2) (G.1)

is a martingale (F, P). Thus, we should have E[Mt |Fs ] = Ms or, by (C.4), E Mt L|Fs E [L|Fs ] = Ms .

Using the martingale property (E.1) of Lt , this is the same as E Mt Lt |Fs = Ms Ls i.e. Mt Lt should be a martingale (F, P). Since d(Mt Lt ) = (t dt)Lt + Mt (et 1)Lt (t dt) + (Mt + 1)Lt et Mt Lt ) dNt = Lt dt t + et t + (Mt + 1)Lt et Mt Lt ) dMt . 47

 we conclude that the martingale property is obtained by choosing t = ln t ln t . The multivariate case goes in the same way; just replace by vector-valued processes.

Some useful vector and matrix results

Vectors and matrices are denoted by in bold letters, lower and upper case, respectively. They may be equipped with topscripts indicating dimensions, e.g. Anm has n rows and m columns. We may write A = (ajk )kK to emphasize jJ the ranges of the row index j and the column index k. The transpose of A is denoted by A . Vectors are invariably taken to be of column type, hence row vectors appear as transposed. The identity matrix is denoted by I, the vector with all entries equal to 1 is denoted by 1, and the vector with all entries equal to 0 is denoted by 0. By Dj=1,...,n (aj ), or just D(a), is meant the diagonal matrix with the entries of a = (a1 , . . . , an ) down the principal diagonal. The ndimensional Euclidean space is denoted by Rn , and the linear subspace spanned by the columns of Anm is denoted by R(A). A diagonalizable square matrix Ann can be represented as
n

A = Dj=1,...,n (j ) 1 =
j=1

j j j ,

(H.1)

where the j are the columns of nn and the j are the rows of 1 . The j are the eigenvalues of A, and j and j are the corresponding right and left eigenvectors, respectively. Eigenvectors (right or left) corresponding to eigenvalues that are distinguishable and non-null are mutually orthogonal. These results can be looked up in e.g. [20]. The exponential function of Ann is the n n matrix dened by

exp(A) =
p=0

1 p A = Dj=1,...,n (ej ) 1 = p!

e j j j ,
j=1

(H.2)

where the last two expressions follow from (H.1). The matrix exp(A) has full rank. If nn is positive denite symmetric, then 1 , 2 = 1 2 denes an inner product on Rn . The corresponding norm is given by = , 1/2 . If Fnm has full rank m ( n), then the , -projection of onto R(F) is F = PF , where the projection matrix (or projector) PF is PF = F(F F)1 F . The projection of onto the orthogonal complement R(F) is F = F = (I PF ) . 48 (H.4) (H.3)

Its squared length, which is the squared , F

2

-distance from to R(F), is (H.5)

= (I PF ) .

The cardinality of a set Y is denoted by |Y|. For a nite set it is just its number of elements.

Ragnar Norberg Department of Statistics London School of Economics Houghton Street London WC2A 2AE Tel: 020-7955-6030 Fax: 020-7955-7416 e-mail: R.Norberg@lse.ac.uk

49