Pricing risk
An actuary puts a defensible price on uncertain future cash flows. They are the original risk data-scientists: given mortality tables, claim histories, and the cost of money, they answer the only question an insurer, a pension fund, or a bank ever really asks — how much should this promise cost today, and how much capital must we hold so the promise survives a bad year? The central act has two halves that never come apart: you price the risk, then you hold capital against it. The price is the mean of an uncertain bill; the capital is your armor against its tail.
The applied playground already taught the engine — expected utility, the break-even probability $p^{\ast}=\dfrac{C}{L-D}$, and the value of information. But every problem there was a single one-period bet: one premium, one loss, one decision, all settled by next year. Real liabilities run for decades and arrive as distributions, not point losses. Three tools lift the actuary above that one-period world, and this page is built on them.
A dollar promised in twenty years is not a dollar. With interest rate $i$ per period, the discount factor is $$v=\frac{1}{1+i},$$ and a future cash flow is worth $v^{t}$ times its face value today. Every premium and every claim in this track gets pulled back to the present through $v^{t}$ before anything is compared. This is the financial mathematics the playground's bets quietly skipped.
The playground lost a fixed $L$ with probability $p$. An actuary instead carries the entire distribution of the bill, decomposed as frequency $\times$ severity: how many claims arrive, and how large each one is. The mean sets the price; the tail — the rare, ruinous outcome far from the average — sets the capital. Mortality enters the same way, as the survival and death probabilities ${}_{k}p_x$ (chance a life aged $x$ survives $k$ more years) and $q_x$ (chance it dies within the year). This is the probability the playground compressed into a single number.
A premium equal to the expected loss is fair but suicidal: half the time you lose, and there is no margin for the bad years or for holding capital. So actuaries add a loading on top of the pure expected cost, and markets go further with a change of measure — pricing not under the real-world probabilities $\mathbb{P}$ (what actually happens) but under a risk-neutral measure $\mathbb{Q}$ (what prices behave as if they believe). The gap between $\mathbb{P}$ and $\mathbb{Q}$ is the price of risk itself, and it is what separates a fair premium from a market one.
The keystone that ties the three together is the equivalence principle: set the present value of premiums in equal to the present value of benefits out. It is the grown-up form of $p^{\ast}=C/(L-D)$ — the same break-even logic, but discounted across time and weighted by survival rather than balanced in a single period: $$\mathrm{APV}(\text{premiums})=\mathrm{APV}(\text{benefits}).$$ Here $\mathrm{APV}$ is the actuarial present value — a cash flow discounted by $v^{t}$ and weighted by the probability it is actually paid. Two building blocks recur throughout: $A_x$, the APV of a \$1 death benefit on a life aged $x$, and $\ddot{a}_x$, the APV of a \$1 annuity-due paid at the start of each year that life survives. Premiums are just the level payment that makes the two sides balance.
Same field, harder questions. The playground asked whether a single \$139.99 premium was worth it this year; here we ask what a 30-year promise is worth today, and what cushion keeps it solvent through the bad tail. Prerequisites are the applied playground (expected utility, break-even probability, the value of information) and, behind it, the decision-theory ladder (Bayes nets → utility → MEU → VPI). If $p^{\ast}=C/(L-D)$ already feels obvious, you have the one prerequisite that matters; everything below adds time, the full distribution, and the price of risk on top of it.
This is a problem track, not an essay, so every thread follows the same seven beats — Setup (the scenario and what to price), Givens & assumptions (the interest rate, the mortality or claim model, the loadings on the table), From first principles (where the formula comes from, derived rather than quoted), Worked numbers (the arithmetic, step by step), Answer (the boxed price or capital), In practice (how a working actuary actually does this — the regulation, the margin, the caveat the clean math hides), and Takeaway (the transferable lesson). It is the playground's five beats with two added: a first-principles derivation up front, so no formula arrives by decree, and a practitioner's note at the end, so the textbook answer meets the real desk.
1 · The price of a life — the equivalence principle present value × survival
An applicant aged $x$ wants a small 3-year term policy: if they die within the next three years, their family receives $\$1000$ at the end of that year; if they survive, nothing is paid. They pay a single level premium each year they are alive. What yearly premium is fair — neither a gift to the policyholder nor a hidden tax on them?
Benefit $B=\$1000$, paid at the end of the year of death. Interest $i=5\%$, so the one-year discount factor is $v=\frac{1}{1.05}$. One-year death probabilities $q_x=0.02,\ q_{x+1}=0.025,\ q_{x+2}=0.03$. Survival to the start of year $k$ is ${}_{k}p_x$: ${}_{0}p_x=1$, ${}_{1}p_x=1-q_x=0.98$, ${}_{2}p_x={}_{1}p_x\,(1-q_{x+1})=0.98\times0.975=0.9555$. The level premium $P$ is paid at the start of each year while the life is alive (an annuity-due). Mortality is assumed independent of interest; no expenses or profit yet (net premium).
A premium is fair when the expected present value (APV) of premiums collected equals the APV of benefits paid — the equivalence principle. This is break-even $p^{\ast}=C/(L-D)$ grown up: instead of a single certain trade, both sides are random cash-flows weighted by the chance the life is still paying or has just died, and discounted to today. Writing $\ddot{a}$ for the premium annuity-due and $A$ for the benefit APV, $$P\cdot\ddot{a} \;=\; A \quad\Longrightarrow\quad P=\frac{A}{\ddot{a}},\qquad A=B\sum_{k=0}^{2} v^{k+1}\,{}_{k}p_x\,q_{x+k},\qquad \ddot{a}=\sum_{k=0}^{2} v^{k}\,{}_{k}p_x.$$
$\phantom{A}=1000\,[\,0.019048+0.022222+0.024763\,]=\$66.03$.
$\phantom{\ddot{a}}=1 + 0.93333 + 0.86667 = 2.8000$. The first $\$1$ is certain (paid today, alive); later premiums are scaled by survival and discounted.
The fair net premium is $\approx\$23.58$ per year per $\$1000$ of 3-year term cover — the level yearly payment whose expected discounted total exactly funds the expected discounted death benefit.
This is the net premium. The price actually charged (the gross premium) adds loadings for expenses, profit, and adverse-deviation risk — the subject of the next thread. Real pricing replaces these illustrative $q$'s with a published mortality table and a prudent (conservative) interest basis rather than a single best-estimate rate. And once the policy is in force, the same two APVs define the reserve the insurer must hold: $\text{reserve}=\text{APV(future benefits)}-\text{APV(future premiums)}$, which starts near zero and grows as the cheap early years pre-fund the riskier later ones.
Every insurance price is a break-even in disguise, but here the "equals" is taken in expected present value across an uncertain lifetime: discount each future dollar, weight it by the probability the cash-flow actually happens, and set the two streams equal. The engine that solves $p^{\ast}=C/(L-D)$ for a one-shot bet is the same machine — just summed over years and survival.
2 · From net premium to the price you pay loadings & premium principles
Thread 1 ended with a net premium of $P_{\text{net}}=\$23.58$ — the bare expected cost of claims. But no insurer can sell at cost: it carries salaries, commissions, capital against a bad year, and shareholders to pay. So what does the customer actually pay, and how is that markup set in a principled way rather than guessed?
Net premium $P_{\text{net}}=E[S]=\$23.58$, where $S$ is the random claim cost. Combined loading $\theta=0.20$ (20%), bundling expenses, profit, and a risk margin. Gross (office) premium $G$ is the posted price. We assume the loading is applied multiplicatively as an expense ratio on the gross — the standard "expense-as-fraction-of-premium" convention — so $G(1-\theta)$ must cover claims. Alternative premium principles reference $E[S]$ and $\mathrm{SD}[S]$; $\alpha$ is a safety multiplier and $a>0$ the insurer's risk-aversion coefficient.
Of every premium dollar collected, a fraction $\theta$ is consumed by expenses, profit, and margin before any claim is paid; the surviving fraction $(1-\theta)$ must equal the expected claims. Setting "what survives" equal to "what claims cost," $G\,(1-\theta)=E[S]$, gives the gross-up rule $$G=\frac{P_{\text{net}}}{1-\theta}=\frac{E[S]}{1-\theta}.$$ This is the expense-ratio special case of a general loaded principle $H[S]=(1+\theta)E[S]$ (expected-value), or $H[S]=E[S]+\alpha\,\mathrm{SD}[S]$ (standard-deviation), or the exponential/utility premium $H[S]=\tfrac1a\ln E\!\left[e^{aS}\right]$ — the law-thread's expected utility, now wielded by the seller.
The customer pays a gross premium of $\boxed{\$29.48}$ — net claims of \$23.58 plus \$5.90 of loading — and the price implies a claim rate $1/0.80=1.25\times$ the actuary's true rate.
A working actuary splits that 20% into named pieces: acquisition commission, per-policy maintenance, the cost of capital held against adverse deviation, plus margins for anti-selection and lapse risk. Loadings can be flat (per-policy) and proportional, not a single multiplier; regulators cap them (minimum loss ratios, rate-filing review) so the markup cannot run away. The exponential principle is the rigorous link: a larger risk-aversion $a$ or a thinner capital base pushes $\tfrac1a\ln E[e^{aS}]$ above $(1+\theta)E[S]$, so the loading is the price of the insurer's own risk appetite.
A posted price reveals a probability plus the seller's margin: divide by the loss ratio to strip the loading and recover the true rate. This is the seller-side mirror of the playground's slogan — the same utility-and-break-even engine the law thread used to decide whether to settle, now run by the party setting the price.
3 · How much capital keeps the doors open ruin theory
An insurer collects premium continuously and pays claims as they arrive. Hold too little surplus and one bad run of losses ends the company. The board asks a sharp question: how much starting capital \$u must we hold so the probability of ever going insolvent, over an infinite horizon, stays below 1%?
Claims arrive as a Poisson process with rate $\lambda$ (claims per unit time) and i.i.d. sizes $X$ that are exponential with mean $\mu = \$1000$. Security loading $\theta = 0.2$, so premium income arrives at rate $c = (1+\theta)\lambda\mu$. Surplus starts at $u$ and follows $U(t) = u + c\,t - S(t)$, where $S(t)$ is total claims paid by time $t$. Ruin is the event that $U(t) < 0$ for some $t$; its probability is $\psi(u)$. Target: $\psi(u) \le 0.01$.
Premium must beat expected claims or ruin is certain: the net drift per unit time is $c - \lambda\mu = \theta\lambda\mu > 0$, so the loading is what keeps the path climbing on average. The adjustment coefficient $R$ is the positive root that balances the moment-generating forces of premium inflow against claim outflow; for exponential claims it solves cleanly to $R = \theta / [(1+\theta)\mu]$. The Cramer-Lundberg theorem then gives the ruin probability exactly for exponential claims: $$\psi(u) = \frac{1}{1+\theta}\,e^{-R u}.$$
About \$26,540 of surplus caps the lifetime probability of ruin at 1%; equivalently it is roughly 26.5 mean-claim payouts held in reserve, the buffer that absorbs an unlucky early cluster of losses before premium income catches up.
The clean closed form holds only because claims are exponential; with heavy-tailed (e.g. lognormal or Pareto) losses there is no exact $\psi(u)$, and the most you get analytically is Lundberg's general inequality $\psi(u) \le e^{-R u}$, an upper bound that drops the leading $\tfrac{1}{1+\theta}$ factor. Modern solvency regimes also abandon the infinite-horizon view entirely: Solvency II and US RBC instead hold capital against a one-year 99.5% Value-at-Risk (or a Tail-VaR / expected shortfall), computable by simulation for any loss distribution. The cost of holding that regulatory capital is itself a loading folded back into the premium (see thread 2).
Capital is the insurer self-insuring against its own tail. Ruin probability decays geometrically in surplus at rate $R$, so the marginal solvency bought per extra dollar shrinks fast, which is why required capital grows only like $\ln(1/\varepsilon)$ as the target $\varepsilon$ tightens. The security loading is the price of solvency, ultimately paid by the customer through the premium, and it buys back their own protection. Dial $\theta$ in the playground engine and watch $R$, hence the required $u$, move.
4 · Two probabilities: real-world vs the market's P vs Q valuation
A catastrophe bond pays a healthy coupon, but if a hurricane above a defined size strikes you lose your principal. Your cat model says the annual trigger probability is about $1\%$, and a trigger wipes the principal, so the modelled expected loss is also about $1\%$. Yet the bond trades at a spread of about $4\%$ over the risk-free rate, four times the loss you modelled. Is the market simply wrong?
Real-world (physical) measure $P$; risk-neutral (market-consistent) measure $Q$. Modelled annual trigger probability $p_P \approx 0.01$. Loss-given-trigger $\text{LGT}\approx 100\%$, so the real-world expected loss is $\text{EL}_P = p_P \times \text{LGT} \approx 1\%$ of principal. Market spread over risk-free $s \approx 4\%$ per year. One-year bond; ignore discounting and recovery beyond $\text{LGT}$ for the toy model. The catastrophe is the kind of rare, correlated, hard-to-hedge tail event investors dislike.
There are two probabilities, not one. Under $P$ the actuary's model gives the real-world expected loss, $\text{EL}_P = p_P\,\text{LGT}$. But price is not set under $P$; it is set under the market-consistent measure $Q$, where $\text{price}=$ expected discounted payoff. Investors who hold an undiversifiable catastrophe demand extra reward for the tail, so they pay less than the $P$-fair value; equivalently the embedded loss probability is inflated. With the spread compensating for the risk-neutral expected loss, $s \approx \text{EL}_Q = p_Q\,\text{LGT}$, the quoted price implies a $Q$-probability above the modelled $P$-probability, $$ p_Q \;=\; \frac{s}{\text{LGT}} \;\ge\; p_P, \qquad \text{multiple} \;=\; \frac{s}{\text{EL}_P}. $$
The market is pricing a $4\%$ implied ($Q$) probability against your modelled $1\%$ ($P$) probability, a $4\times$ multiple, and the $3$-point wedge is compensation for undiversifiable catastrophe risk, not a mistake in the model.
Insurance-linked-securities desks quote exactly these multiples (spread divided by expected loss) to compare deals. A working actuary keeps the measures separate: $P$ for reserving and projecting actual losses, $Q$ for market value and mark-to-market. The wedge widens after big loss years (capital is scarce, so risk is dear) and narrows in soft markets. They also refine the toy model with partial $\text{LGT}$, multi-year and aggregate triggers, reset spreads, and basis risk, none of which collapse the $P$, $Q$ gap.
A quoted price yields a $Q$-probability; your model yields a $P$-probability; the gap between them is risk aversion to the tail. This is the rigorous twin of the betting-line vig from thread 2: there the seller's margin inflated the implied number, here it is the price of risk. For hedgeable assets no-arbitrage forces $P$ and $Q$ to reconcile, but for catastrophe risk, which cannot be replicated, they stay apart, and that distance is the premium.
5 · Trust your data, or the manual? credibility (Bayesian shrinkage)
A trucking fleet's own claims have averaged \$300 per vehicle-year over the last three years, but the manual (book) rate for its risk class is \$500. What do you charge the fleet next year: its own experience, the book rate, or some blend of the two? Lean entirely on three years of one fleet's luck and you over-react to noise; ignore that experience and you waste a real signal.
Fleet's own average loss $\bar{X}=\$300$ per vehicle-year, observed over $n=3$ exposure-years. Class manual rate (the prior mean) $\mu=\$500$. The credibility constant $k=\dfrac{\text{E[process variance]}}{\text{Var[hypothetical means]}}=7$, estimated once from the whole portfolio's variance components. We seek a premium estimate $\hat{P}$ that is the best linear blend of $\bar{X}$ and $\mu$.
One economic principle: weight each source by how much you trust it. With little data the fleet's mean is mostly noise, so lean on the book; with much data the fleet's own record swamps the noise, so lean on the fleet. Bühlmann credibility makes that weight the best linear estimator: put weight $Z$ on experience and $1-Z$ on the book, where $Z$ rises with exposures $n$ and falls with the relative noise $k$, $$Z=\frac{n}{n+k},\qquad \hat{P}=Z\,\bar{X}+(1-Z)\,\mu.$$
Charge $\hat{P}=\$440$. The estimate is shrunk 30% of the way from the book's \$500 toward the fleet's own \$300. That is exactly as far as three years of data earns the right to move it.
The constant $k$ is not guessed; it comes from variance components estimated across the whole portfolio. Here $k$ is expected within-risk (process) variance over between-risk (hypothetical-means) variance, so homogeneous classes get small $k$ and high credibility, noisy ones get large $k$. This same machinery drives experience rating, bonus-malus / no-claims-discount schemes in motor insurance, and workers' compensation. Bühlmann's linear blend is the best linear approximation to the full Bayesian posterior mean, and matches it exactly when losses are conjugate (e.g. Poisson-Gamma).
Credibility is Bayesian shrinkage with a dollar tag. Each extra exposure-year raises $Z$ toward $1$ and pulls the premium toward the fleet's own truth, so data has a measurable price: it is the seller's cousin of the medicine problem's VPI. More data sharpens the estimate, and that sharpening is an actuarial value of information.