Mixture and Spliced Distributions for Insurance

Why Use Mixture and Spliced Distributions

Standard parametric distributions sometimes fail to capture the complexity of real insurance loss data. A single distribution may fit the body of the data well but miss the tail, or it may accommodate the tail at the expense of the body. Mixture and spliced distributions address this by combining multiple distributional components, providing the flexibility to model different features of the loss distribution with different parametric forms.

These models arise naturally when the insured population is heterogeneous. An auto insurance portfolio may contain both cautious drivers (generating small, infrequent claims) and reckless drivers (generating larger claims). A single exponential distribution cannot capture both groups well, but a mixture of two exponentials can. Similarly, property losses may follow one pattern for routine damage and a very different pattern for catastrophic events, motivating a spliced model with different distributions for different loss ranges.

Mixture Distributions

A k-point mixture combines k distributions with weights w_1, w_2, ..., w_k that sum to 1. The PDF is f(x) = w_1*f_1(x) + w_2*f_2(x) + ... + w_k*f_k(x), where each f_j is a proper density and w_j represents the probability that an observation comes from component j. The CDF is similarly F(x) = w_1*F_1(x) + w_2*F_2(x) + ... + w_k*F_k(x).

The mean of a mixture is the weighted average of the component means: E[X] = w_1*E[X_1] + w_2*E[X_2] + ... + w_k*E[X_k]. The variance is more complex because it includes both within-component and between-component variation: Var(X) = sum of w_j*Var(X_j) + sum of w_j*(E[X_j])^2 - (E[X])^2. This can be rewritten using the law of total variance: Var(X) = E[Var(X|J)] + Var(E[X|J]), where J indicates the mixture component.

Mixture Distribution Examples

A common two-point mixture for insurance losses combines two exponential distributions: f(x) = w*(1/theta_1)*exp(-x/theta_1) + (1-w)*(1/theta_2)*exp(-x/theta_2). If w = 0.7, theta_1 = 500, and theta_2 = 5000, this models a population where 70% of losses come from a low-severity component (mean 500) and 30% from a high-severity component (mean 5000). The overall mean is 0.7*500 + 0.3*5000 = 350 + 1500 = 1850.

The survival function of this mixture is S(x) = 0.7*exp(-x/500) + 0.3*exp(-x/5000). For small x, both components contribute; for large x, the heavy component dominates. This produces a heavier tail than either component alone, which is a useful feature for modeling loss distributions with occasional large claims.

To compute probabilities, simply apply the mixture formula. For instance, P(X > 2000) = 0.7*exp(-2000/500) + 0.3*exp(-2000/5000) = 0.7*exp(-4) + 0.3*exp(-0.4) = 0.7*0.0183 + 0.3*0.6703 = 0.0128 + 0.2011 = 0.2139. Notice how the high-severity component contributes most of the tail probability.

Spliced Distributions

Spliced distributions use different parametric models for different ranges of the loss variable, joined at one or more threshold points. Unlike mixtures, where all components are defined over the entire support and blended by weights, spliced models assign each component to a specific region of the loss range. This gives direct control over the behavior in each region.

For a two-piece spliced model with threshold c, the PDF is: f(x) = alpha * f_1(x) / F_1(c) for 0 < x <= c, and f(x) = (1-alpha) * f_2(x) / S_2(c) for x > c. Here alpha is the probability that a loss falls below c, f_1 is the density for the lower piece (normalized to integrate to alpha over (0, c]), and f_2 is the density for the upper piece (normalized to integrate to 1-alpha over (c, infinity)). Each component f_j is a proper density on its natural domain, and the normalizing factors F_1(c) and S_2(c) ensure the spliced density integrates to 1.

A common spliced model uses an exponential for losses below c and a Pareto for losses above c. The exponential handles the routine, moderate losses while the Pareto accommodates the heavy tail. This combination provides far more flexibility than either distribution alone.

Computing Moments for Mixtures and Spliced Models

For mixture distributions, the k-th raw moment is E[X^k] = sum of w_j * E[X_j^k]. This follows directly from the linearity of expectation applied to the mixture density. Limited expected values are also weighted averages: E[X wedge d] = sum of w_j * E[X_j wedge d], where each component's limited expected value uses its own distribution.

For spliced distributions, moments require integrating over each piece separately. The mean is E[X] = alpha * E_1[X | X <= c] + (1-alpha) * E_2[X | X > c], where each conditional expectation uses the appropriate component's distribution. The limited expected value E[X wedge d] depends on whether d falls below or above the splice point c. If d <= c, only the lower component matters: E[X wedge d] = alpha * E_1[X wedge d] / F_1(c). If d > c, both components contribute.

Fitting Mixture and Spliced Models

Parameter estimation for mixture distributions uses maximum likelihood or the EM (expectation-maximization) algorithm. The EM algorithm iterates between computing posterior probabilities that each observation belongs to each component (E-step) and updating component parameters and weights (M-step). This is particularly useful when the number of components is known but component membership is unobserved.

For spliced models, the threshold c may be fixed by the analyst (based on expert judgment or visual inspection of the data) or estimated as part of the fitting procedure. With c fixed, the parameters of each component are estimated from the observations in the corresponding range. Maximum likelihood estimates account for the truncation of each component to its designated range.

Common Pitfalls

A frequent error with mixture distributions is computing the variance as the weighted average of component variances alone, forgetting the between-component variance term. The correct formula is Var(X) = sum of w_j*(Var(X_j) + (E[X_j])^2) - (sum of w_j*E[X_j])^2. Always use the law of total variance or compute E[X^2] - (E[X])^2 from the mixture moments.

For spliced distributions, students often forget the normalizing constants that ensure each piece integrates to the correct weight over its designated range. Without these constants, the spliced density does not integrate to 1 and all probability calculations will be incorrect. Double-check that alpha * integral of the lower piece + (1-alpha) * integral of the upper piece equals 1.

Study Tips

Exam STAM tests your ability to compute probabilities, moments, and limited expected values from mixture and spliced distributions. Practice with two-component mixtures of exponentials (the most common exam setup) and with exponential-Pareto spliced models. For each problem, identify whether you have a mixture (components blended by weight over the full range) or a spliced model (components assigned to specific ranges). Work through the moment calculations carefully, paying attention to normalizing constants and conditional expectations. Building familiarity with these models will also help with the broader topic of model selection, where mixtures and spliced distributions compete with single parametric families for best fit.