Why Use a Saddlepoint Approximation?
A normal approximation can be useful for an aggregate loss near its mean, but it may be inaccurate in a tail. A saddlepoint approximation improves the description of a tail by choosing an exponential tilt that makes the threshold of interest the mean under a changed probability distribution. This is especially useful when an aggregate distribution is difficult to calculate directly.
Let \(T=\sum_{j=1}^{N}X_j\) be an aggregate loss, where \(N\) is the claim count and the \(X_j\) are independent, identically distributed claim severities, independent of \(N\). The cumulant generating function (CGF) of \(T\) is \(K(t)=\log E[e^{tT}]\), for values of \(t\) where this expectation is finite. Its first two derivatives are \(K'(t)=E[T]\) and \(K''(t)=\operatorname{Var}(T)\) at \(t=0\); more generally, \(K'(t)\) and \(K''(t)\) describe the mean and variance under exponential tilting.
Compound Poisson CGF and the Tail Formula
For a compound Poisson model, \(N\sim\operatorname{Poisson}(\lambda)\), with severity moment generating function \(M_X(t)=E[e^{tX}]\). The aggregate CGF follows from conditioning on \(N\):
This expression applies only where \(M_X(t)\) is finite. The saddlepoint equation is therefore \(\lambda M_X'(t)=x\), with the additional requirement that \(t\) be in the CGF domain. For a positive severity model, \(K'\) is increasing wherever \(K''>0\), so there can be at most one root in that domain. The sign of the root is informative: a threshold above \(E[T]\) generally gives \(t>0\), while a threshold below \(E[T]\) generally gives \(t<0\).
For an interior threshold \(x\), define the signed root \(w\) and the quantity \(u\) as follows. The signed square root preserves whether the threshold is above or below the mean.
For a continuous, nonlattice aggregate and a nonzero saddlepoint, the Lugannani–Rice approximation to the upper tail is
where \(\Phi\) and \(\phi\) are the standard normal cumulative distribution function and density. When \(x\) is below the mean, it is often clearer to calculate the lower tail directly:
At \(x=E[T]\), both \(w\) and \(u\) are zero and these displayed expressions should not be substituted directly; use the appropriate limiting form or a different approximation. The formulas also do not automatically handle atoms or lattice distributions. For example, a compound Poisson aggregate has an atom at zero, and an aggregate of fixed-size claims is lattice-valued.
Why Exponential Tilting Helps
The saddlepoint equation is not just a device for finding a number. It selects a probability distribution under which the rare threshold is centered. Define the tilted distribution by weighting outcomes in proportion to \(e^{tT}\). Its density, when one exists, is \(f_t(y)=e^{ty-K(t)}f(y)\). Under this distribution the mean is \(K'(t)\) and the variance is \(K''(t)\). Choosing \(K'(t)=x\) therefore makes \(x\) the tilted mean.
The change of measure gives the following identity, which also explains why the method can improve a tail approximation:
Here \(E_t\) denotes expectation under the tilted distribution. The exponential factor in the expectation reduces the influence of outcomes far beyond the threshold; under the tilt, the threshold itself is centered rather than far out in a tail. A normal approximation under this changed distribution gives a basic saddlepoint approximation. The Lugannani–Rice expression refines it with a correction involving \(w\) and \(u\), improving the approximation around the threshold. This is an approximation, not an exact distributional identity.
Worked Examples
Worked Example: An Upper Tail for Exponential Severities
A hypothetical portfolio has \(N\sim\operatorname{Poisson}(16)\) claims. Claim severities are independent exponential random variables with rate \(\beta=0.5\) per thousand dollars, so the mean severity is 2 thousand dollars. Find the Lugannani–Rice approximation to \(P(T\ge40)\), where \(T\) is measured in thousands of dollars. There is no discounting or mortality assumption in this aggregate model.
Method and CGF. For an exponential severity, \(M_X(t)=\beta/(\beta-t)\) for \(t<\beta\). Thus \(K(t)=16t/(0.5-t)\), \(K'(t)=8/(0.5-t)^2\), and \(K''(t)=16/(0.5-t)^3\). The aggregate mean is \(K'(0)=32\), so an upper-tail threshold should have a positive root.
Solve and check the root. Setting \(K'(t)=40\) gives \(0.5-t=\sqrt{8/40}=\sqrt{0.2}=0.4472\), hence \(t=0.0528\). This root is positive and satisfies the required domain \(t<0.5\).
Calculate the LR quantities. Using the unrounded root, \(K(t)=1.8885\), \(K''(t)=178.8854\), and \(tx=2.1115\). Therefore \(w=\sqrt{2(2.1115-1.8885)}=0.6677\), and \(u=0.0527864\sqrt{178.8854}=0.7060\).
Approximate the tail. With \(\Phi(0.6677)\approx0.7479\) and \(\phi(0.6677)\approx0.3193\), \(P(T\ge40)\approx1-0.7479+0.3193(1/0.7060-1/0.6677)=0.2262\). The approximation is a probability between zero and one, and the positive root is consistent with a threshold above the mean. Rounding is to four decimal places.
Common mistake avoided. Using \(t=0\) because it is the point at which the CGF is easiest to evaluate does not center the tilted distribution at 40. The saddlepoint equation, not convenience, determines the root.
Worked Example: A Lower Tail and a Negative Root
A hypothetical aggregate has \(N\sim\operatorname{Poisson}(9)\) claims, with independent exponential severities of rate 1 per thousand dollars (mean severity 1 thousand dollars). Approximate \(P(T\le6)\). No discounting or mortality assumption is involved.
Method and root. Here \(K(t)=9t/(1-t)\) for \(t<1\), \(K'(t)=9/(1-t)^2\), and \(K''(t)=18/(1-t)^3\). The aggregate mean is 9, so the threshold 6 is below the mean. Solving \(K'(t)=6\) gives \(1-t=\sqrt{9/6}=1.2247\), so \(t=-0.2247\). This negative root is in the domain \(t<1\).
Calculate the standardized quantities. At this root, \(K(t)=-1.6515\), \(K''(t)=9.7980\), and \(tx=-1.3485\). Hence \(w=-\sqrt{2\{-1.3485-(-1.6515)\}}=-0.7785\), where the negative sign matches \(t\). Also \(u=-0.2247\sqrt{9.7980}=-0.7035\).
Approximate the lower tail. Using \(\Phi(-0.7785)\approx0.2182\) and \(\phi(-0.7785)\approx0.2947\), \(P(T\le6)\approx0.2182+0.2947(1/(-0.7785)-1/(-0.7035))=0.2586\). The negative root correctly identifies a lower-tail calculation. Rounding is to four decimal places.
Common mistake avoided. Replacing the root by its absolute value changes the sign of \(w\) and gives the wrong side of the distribution. Retain the sign of \(t\) when defining the signed root.
Worked Example: Detecting When No Finite Root Exists
A hypothetical portfolio has \(N\sim\operatorname{Poisson}(3)\) claims, with independent exponential severities of rate 2 per thousand dollars. Determine whether a finite saddlepoint can be used at threshold \(x=0\), and calculate \(P(T>0)\). There is no discounting or mortality assumption.
Check existence. The CGF is \(K(t)=3t/(2-t)\) for \(t<2\), and \(K'(t)=6/(2-t)^2\). For every finite \(t<2\), \(K'(t)>0\), so the equation \(K'(t)=0\) has no finite solution. The usual LR formula cannot be applied at this threshold.
Use the model's atom at zero. Since every severity is strictly positive, \(T=0\) exactly when \(N=0\). Therefore \(P(T=0)=P(N=0)=e^{-3}=0.0498\), and \(P(T>0)=1-e^{-3}=0.9502\), to four decimal places.
Common mistake avoided. Treating a limiting root as a finite saddlepoint and substituting it into the LR formula is not valid. Here the exact probability follows directly from the probability of zero claims.
Existence Checks and Exam Traps
Before evaluating a tail approximation, check the model and the root rather than assuming a solution exists. For a nondegenerate aggregate with \(K''(t)>0\), \(K'\) is strictly increasing on the CGF domain. A practical check is to identify the range of \(K'\) on that domain, then verify that \(x\) lies within it. If \(x\) is in the range, strict increase gives at most one root; if it is outside the range, there is no finite root.
For a compound Poisson model, use \(K(t)=\lambda(M_X(t)-1)\) and record the severity MGF's domain.
Compare \(x\) with \(K'(0)=E[T]\). Solve \(K'(t)=x\), then verify the root is in the CGF domain and that \(K''(t)>0\).
For a threshold below the mean, the lower-tail formula can be clearer. Keep the sign of \(t\) in \(w\) and \(u\).
Use the ordinary LR formula only when its smooth, nonlattice assumptions are appropriate. Check that the resulting probability is between zero and one.
A common exam trap is to find an algebraic root but forget that the CGF may be infinite there. Another is to apply a continuous formula to a threshold at an atom or to a lattice-valued aggregate without a suitable adjustment. In particular, the compound Poisson atom at zero should be handled explicitly when the threshold concerns zero claims. If no finite root exists, do not force the LR calculation; use an exact probability when available or a method appropriate to the model.
The method is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. Its central discipline is to check existence, domain, and root sign before calculating a tail approximation.
Check Your Understanding
For each problem, identify the relevant root or applicability check before using the requested result.
- A compound Poisson aggregate has \(\lambda=9\) and exponential severities with rate 1 per thousand dollars. For threshold \(x=12\), find the saddlepoint \(t\) to four decimal places and state whether it is positive or negative. (Answer: \(t=0.1340\), positive.)
- A compound Poisson aggregate has \(\lambda=4\) and exponential severities with rate 2 per thousand dollars. At \(x=1\), find the saddlepoint to four decimal places and verify it is in the CGF domain. (Answer: \(t=-0.8284\), in the domain \(t<2\).)
- For \(K(t)=6(e^{2t}-1)\), solve \(K'(t)=18\) for \(t\), to four decimal places. (Answer: \(t=\tfrac12\ln(1.5)=0.2027\).)
- A compound Poisson aggregate has \(\lambda=5\) and strictly positive exponential severities. At \(x=0\), is there a finite saddlepoint, and what are \(P(T=0)\) and \(P(T>0)\), to four decimal places? (Answer: No finite saddlepoint; \(P(T=0)=0.0067\) and \(P(T>0)=0.9933\).)
- Suppose \(K'\) is strictly increasing on its CGF domain, with range \((0,25)\). For which thresholds \(x\) in that range is there exactly one finite saddlepoint? (Answer: Every \(x\) with \(0<x<25\); there is no finite root for thresholds outside that range.)