Integrated Multi-State Thiele Reserve Workshop

Why an Integrated Reserve Calculation Needs More Than an ODE

A multi-state reserve calculation brings together the model, the cash-flow basis, and the direction of calculation. Benefits and expenses are outgo; premiums are income. Transition benefits are paid only when the specified transition occurs. If a reserve is specified at the end of the coverage period, the Thiele equations must then be integrated backward to the valuation time. A sign or boundary error can produce a numerically tidy answer that does not represent the contract.

This workshop uses reserves conditional on the insured being in each living state. It focuses on constructing and checking a complete calculation, rather than on a particular numerical solver. The methods are aligned with the SOA Exam ALTAM (Advanced Long-Term Actuarial Mathematics) syllabus.

Definition: Let $V_i(t)$ be the prospective reserve at time $t$, conditional on being in state $i$. Let $\delta$ be the force of interest. Let $c_i(t)$ denote the continuous rate of net outgo in state $i$, with benefits and expenses positive and premiums negative. For a transition $i\to j$, let $\mu_{ij}(t)$ be its intensity and $B_{ij}(t)$ the lump sum paid at that transition, with a positive amount denoting outgo. Then, for a living state $i$, $$ V_i'(t)=\delta V_i(t)-c_i(t)+\sum_{j\ne i}\mu_{ij}(t)\bigl(V_i(t)-V_j(t)-B_{ij}(t)\bigr). $$ For a transition to a terminal state, use $V_j(t)=0$ after settlement. This form assumes the stated cash flows occur continuously or at transitions, the intensities and interest basis are specified, and $V_i(t)$ is a prospective net-outgo reserve. Other reserve or cash-flow sign conventions require corresponding changes to the equation.

The transition term is easy to misread. During a short interval, a transition replaces the reserve for state $i$ by the transition benefit plus the reserve in state $j$. The amount by which the conditional future outgo changes is $B_{ij}+V_j-V_i$. Since the Thiele equation is written for the derivative of the current reserve, this change enters with a minus sign, giving $V_i-V_j-B_{ij}$.

Deriving the Transition Term and the Backward Direction

For a short interval of length $dt$, the probability of a transition $i\to j$ is $\mu_{ij}(t)dt+o(dt)$. Conditional on no transition, the reserve at the end of the interval is $V_i(t+dt)$. If the transition occurs, the immediate outgo is $B_{ij}(t)$, followed by reserve $V_j(t+dt)$. Discounting the future amount over the interval gives, to first order,

V_i(t)=c_i(t)dt+e^{-\delta dt} \left[ V_i(t+dt)+\sum_{j\ne i}\mu_{ij}(t)dt \bigl(B_{ij}(t)+V_j(t+dt)-V_i(t+dt)\bigr) \right]+o(dt).

Using $e^{-\delta dt}=1-\delta dt+o(dt)$ and $V_i(t+dt)=V_i(t)+V_i'(t)dt+o(dt)$, subtract $V_i(t)$, divide by $dt$, and let $dt$ tend to zero. This yields the stated equation. A terminal condition such as $V_i(T)=0$ is imposed at the end of coverage; it does not mean the reserve at earlier times is zero.

For numerical work, if $f(t,\mathbf V)$ denotes the vector of Thiele derivatives, a backward Euler step of length $h>0$ is approximately

\mathbf V(t-h)\approx \mathbf V(t)-h f(t,\mathbf V(t)).

The minus sign is essential: the differential equation gives derivatives in increasing calendar time, while the calculation moves from $T$ toward the valuation time.

Worked Example: A One-State Sign and Boundary Check

Consider a hypothetical one-year term benefit. The force of interest is $\delta=0.04$, the death intensity is constant at $\mu=0.10$, the death benefit is \$1,000, and a continuous premium of \$30 per year is received while the insured is alive. There are no other expenses, and the reserve at the end of the year is $V(1)=0$. Find $V(0)$.

Method. The net outgo rate while alive is $c=-30$, because the premium is income. Death is a transition to a terminal state with reserve zero and benefit $B=1{,}000$. Thus

V'(t)=0.04V(t)-(-30)+0.10\bigl(V(t)-1{,}000\bigr) =0.14V(t)-70.

The constant solution is $V=500$, since $0.14(500)-70=0$. Write $W(t)=V(t)-500$. Then $W'(t)=0.14W(t)$, so $W(t)=Ce^{0.14t}$. Applying $V(1)=0$ gives $C=-500e^{-0.14}$, and therefore

V(0)=500\bigl(1-e^{-0.14}\bigr) =500(1-0.869358\ldots)=65.3209\ldots.

The prospective check gives the same result: the expected net outgo rate at time $s$, conditional on being alive at time zero, is $0.10(1{,}000)-30=70$. The probability of still being alive and the discount factor together give $e^{-0.14s}$. Hence $V(0)=\int_0^1 70e^{-0.14s}\,ds=500(1-e^{-0.14})$. Rounded to the nearest cent, the reserve is \$65.32.

Common mistake avoided: treating the premium as positive outgo would change $-30$ to $+30$. The prospective check makes the error visible: the correct net outgo rate is \$70 per year, not \$130.

Worked Multi-State Calculation with Time-Varying Intensities

Consider a hypothetical product with active state $A$, care state $C$, and terminal death state $D$. Coverage ends at $T=1$, and $V_A(1)=V_C(1)=0$. The force of interest is $\delta=0.04$. In state $A$, a continuous premium of \$10 is received and a continuous expense of \$5 is paid, giving $c_A=5-10=-5$. In state $C$, there is no premium and the continuous expense is \$12, so $c_C=12$.

The transition intensities, in units per year, are $\mu_{AC}(t)=0.10+0.02t$, $\mu_{AD}(t)=0.04$, and $\mu_{CD}(t)=0.08-0.02t$. A transition from $A$ to $C$ pays \$20, a transition from $A$ to death pays \$1,000, and a transition from $C$ to death pays \$300. There are no other transitions. Substitution into the Thiele equations gives

\begin{aligned} V_A'(t) &=0.04V_A(t)+5 +(0.10+0.02t)\bigl(V_A(t)-V_C(t)-20\bigr)\\ &\quad+0.04\bigl(V_A(t)-1{,}000\bigr),\\ V_C'(t) &=0.04V_C(t)-12+(0.08-0.02t)\bigl(V_C(t)-300\bigr). \end{aligned}

The nonzero terminal benefits appear inside the transition terms; they are not continuous benefit rates. To estimate the reserves at time zero, use one backward midpoint step of length one year. This is an approximation, not an exact solution. First evaluate the derivatives at $t=1$, where both reserves are zero:

\begin{aligned} f_A(1,\mathbf 0)&=5+0.12(0-0-20)+0.04(0-1{,}000)=-37.4,\\ f_C(1,\mathbf 0)&=-12+0.06(0-300)=-30. \end{aligned}

The midpoint state estimate is the terminal vector minus one-half of these derivatives, because integration proceeds backward:

\begin{aligned} V_A(0.5)&\approx 0-\tfrac12(-37.4)=18.7,\\ V_C(0.5)&\approx 0-\tfrac12(-30)=15. \end{aligned}

At $t=0.5$, the intensities are $\mu_{AC}=0.11$ and $\mu_{CD}=0.07$. Substituting the midpoint reserves into the equations gives

\begin{aligned} f_A(0.5,\mathbf V_{\mathrm{mid}}) &=0.04(18.7)+5+0.11(18.7-15-20)+0.04(18.7-1{,}000)\\ &=0.748+5-1.793-39.252=-35.297,\\ f_C(0.5,\mathbf V_{\mathrm{mid}}) &=0.04(15)-12+0.07(15-300)\\ &=0.6-12-19.95=-31.35. \end{aligned}

The backward midpoint update is therefore $V_i(0)\approx V_i(1)-f_i(0.5,\mathbf V_{\mathrm{mid}})$. It gives $V_A(0)\approx35.297$ and $V_C(0)\approx31.35$, or \$35.30 and \$31.35 to the nearest cent. As a numerical consistency check, the changes from the terminal values are $35.297-0=35.297=-(-35.297)$ and $31.35-0=31.35=-(-31.35)$, exactly the backward midpoint increments. A finer time mesh is appropriate if greater accuracy is required.

Common mistake avoided: using the intensity at $t=1$ for the whole step would ignore the stipulated time variation. The midpoint method evaluates the transition rates at $t=0.5$ for the reserve update.

Runge–Kutta Stages When Integrating Backward

A higher-order method improves the local approximation, but does not change the direction of integration. For a step from $t$ to $t-h$, where $h>0$, define $k_r=h f$ at the relevant stage. The Runge–Kutta fourth-order (RK4) stage states are formed by subtracting these increments:

\begin{aligned} k_1&=h f(t,\mathbf V),\\ k_2&=h f(t-h/2,\mathbf V-k_1/2),\\ k_3&=h f(t-h/2,\mathbf V-k_2/2),\\ k_4&=h f(t-h,\mathbf V-k_3),\\ \mathbf V(t-h)&\approx \mathbf V(t)-\frac{k_1+2k_2+2k_3+k_4}{6}. \end{aligned}

The next calculation isolates the stage arithmetic. Suppose a hypothetical two-reserve system is being integrated from $t=1$ to $t=0$, with current reserve vector $(500,300)$, $h=1$, and stage increments returned by its time-dependent Thiele vector field as follows: $k_1=(128.54,100.22)$, $k_2=(128.54,100.22)$, $k_3=(127.5862,101.5444)$, and $k_4=(127.5862,101.5444)$. These increments are inputs from the field evaluations; the task here is to apply the backward RK4 update correctly.

Worked Example: Correcting the Backward RK4 Stage States

Method. Subtract half the relevant increment for the second and third stages, and subtract the full third-stage increment for the fourth stage. The second stage is at $t=0.5$:

(500,300)-\frac{k_1}{2} =(500-64.27,\;300-50.11) =(435.73,\;249.89).

The third stage is also at $t=0.5$, but uses $k_2$. Since $k_2=(128.54,100.22)$, its state is likewise

(500,300)-\frac{k_2}{2} =(500-64.27,\;300-50.11) =(435.73,\;249.89).

The fourth stage is at $t=0$. It uses the full $k_3$, so

(500,300)-k_3 =(500-127.5862,\;300-101.5444) =(372.4138,\;198.4556).

Finally, the weighted RK4 increments are

\begin{aligned} \Delta V_1 &=\frac{128.54+2(128.54)+2(127.5862)+127.5862}{6} =\frac{768.3786}{6}=128.0631,\\ \Delta V_2 &=\frac{100.22+2(100.22)+2(101.5444)+101.5444}{6} =\frac{605.2932}{6}=100.8822. \end{aligned}

Thus the estimates at time zero are $(500-128.0631,\;300-100.8822)=(371.9369,\;199.1178)$, or $(\$371.94,\$199.12)$ to the nearest cent. The stage states and the final update both move downward from the current vector because the increments are positive and the integration is backward.

Common mistake avoided: adding the stage increments to the current reserves. For example, the incorrect second-stage state would be $(564.27,350.11)$. The correct calculation is $(500,300)-k_1/2=(435.73,249.89)$; the same subtraction gives the third-stage state. The fourth-stage state is $(372.4138,198.4556)$, not $(627.5862,401.5444)$.

Practical Checks Before Accepting a Reserve Path

An integrated calculation should be checked at more than the final line. The following checks are useful whether the system is solved analytically or numerically.

Check the state and cash-flow definitions.
Confirm that each reserve is conditional on the stated state, premiums are income, and expenses and benefits are outgo. Keep transition lump sums separate from continuous rates.

Check the terminal condition.
Use the stipulated reserve after the final settlement. If coverage ends with all obligations discharged, that may be zero; it is not automatically zero if a terminal payment or continuation obligation remains.

Check backward increments.
For a step from $t$ to $t-h$, use the negative time direction. RK4 stage states and the final update must reflect that direction consistently.

Check the result against the model.
Substitute a reserve and time into the differential equation, check a local cash-flow balance, or repeat the calculation with a finer step. A stable result should also respond sensibly to changes in benefits, premiums, and transition intensities.

These checks do not guarantee that every numerical approximation is sufficiently accurate. In particular, a coarse step can miss rapid changes in an intensity or cash flow. They do identify many avoidable errors: wrong transition benefits, inconsistent premium signs, using rates at the wrong time, or reversing a backward step.

Key takeaway: A reliable multi-state Thiele calculation begins with an explicit cash-flow convention and a correct terminal condition. Write the state-specific equations first, then integrate backward with the time direction shown in every step, and finish with a numerical or prospective consistency check.

Check Your Understanding

For each problem, state the sign convention you use and check whether the requested calculation is forward or backward in time.

A one-year single-state contract has $\delta=0.03$, death intensity $\mu=0.07$, death benefit \$800, and continuous premium income \$20. There are no expenses and $V(1)=0$. Find $V(0)$ using the Thiele equation. (Answer: $V(0)=360(1-e^{-0.10})\approx\$34.26$.)
At a particular time, $V_A=100$, $V_C=40$, $\delta=0.05$, $c_A=-8$, $\mu_{AC}=0.10$, and the $A\to C$ transition benefit is \$10. Also $\mu_{AD}=0.02$, the $A\to D$ benefit is \$500, and $V_D=0$. Find $V_A'$. (Answer: $5+8+0.10(100-40-10)+0.02(100-500)=10$ per year.)
For a scalar reserve equation $V'=0.20V-10$, use one backward Euler step of length $0.25$ from $t=1$, where $V(1)=0$, to estimate $V(0.75)$. (Answer: $V(0.75)\approx2.50$.)
A backward RK4 step starts from $(240,160)$ with $k_1=(20,12)$ and $k_2=(18,10)$. What state vector is used to evaluate the third stage? (Answer: $(240,160)-k_2/2=(231,155)$.)
A backward RK4 step has $k_1=(6,4)$, $k_2=(5,3)$, $k_3=(4,2)$, and $k_4=(3,1)$, starting from $(100,50)$. Find the updated vector. (Answer: $(100,50)-(27/6,15/6)=(95.50,47.50)$.)

What You'll Learn