Modeling and estimating skewed and heavy-tailed populations via unsupervised mixture models

Technical Summary: Modeling and Estimating Skewed and Heavy-Tailed Populations via Unsupervised Mixture Models

Problem Statement
In fields ranging from actuarial science to macroeconomics, population distributions are frequently characterized by skewness and heavy tails. A common challenge is that no single probability model adequately describes both the "body" (bulk) of the data and its extreme tail. Classical size distributions (e.g., Lognormal) often fit the body well but decay too rapidly to capture extreme observations, while Generalized Pareto Distributions (GPD) are theoretically sound for tails but often fail to model the bulk of the data. Existing solutions typically involve "splicing" distributions at a specific threshold or using dynamic mixtures. However, spliced models require the selection of a threshold and impose continuity and differentiability constraints that reduce the number of free parameters. Dynamic mixtures, such as the one proposed by Frigessi et al. (2002), offer smooth densities without thresholds but suffer from intractable likelihoods due to non-closed-form normalizing constants, complicating estimation.

Methodology
The authors propose a static, unsupervised mixture model combining a Lognormal distribution (for the body) and a zero-location Generalized Pareto Distribution (GPD) (for the tail). The probability density function is defined as:
$$f(x; \theta) = p f_1(x; \mu, \sigma^2) + (1-p) f_2(x; 0, \xi, \beta)$$
where $p \in (0,1)$ is the mixing weight, $f_1$ is the Lognormal density, and $f_2$ is the GPD density with location 0.

Key methodological features include:

• Unsupervised Nature: Unlike spliced models, this approach does not require a pre-specified threshold. The transition between the body and tail is governed by the mixing weight $p$, and the density is automatically continuous and differentiable on $[0, +\infty)$ (provided $\xi > -1$).
• Estimation via EM Algorithm: The authors develop an Expectation-Maximization (EM) algorithm for Maximum Likelihood Estimation (MLE).
  • E-step: Computes the posterior probability that an observation belongs to the Lognormal component.
  • M-step: Updates the Lognormal parameters ($p, \mu, \sigma^2$) in closed form. The GPD parameters ($\xi, \beta$) are updated via numerical maximization of a weighted log-likelihood.
• Implementation: The method is implemented in the R package lognGPD.

Key Contributions

1. Novel Model Structure: The paper introduces a static mixture of Lognormal and GPD components. To the authors' knowledge, this specific static configuration (with constant weights) had not been previously considered in the literature.
2. Computational Efficiency: By avoiding the intractable normalizing constant found in dynamic mixtures (Case 2) and the complex constraints of spliced models (Case 1), the proposed model allows for a computationally lighter estimation procedure.
3. Robustness and Flexibility: The model is designed to be flexible enough to fit data generated from various processes, including those where the true distribution differs from the assumed mixture.

Results
The paper evaluates the model through simulation experiments and two real-data applications (automobile insurance claims and operational risk losses).

• Simulation Experiments:

  • Correctly-Specified Setup: When data are simulated from the proposed mixture, the EM algorithm produces consistent estimators. Performance improves significantly with sample size ($n=500$ vs. $n=100$), though GPD parameter estimates ($\xi, \beta$) can be unstable in small samples if the mixing weight $p$ is high (few tail observations).
  • Mis-Specified Setups: The model was fitted to data generated from (1) a composite Lognormal-Pareto distribution (Scollnik, 2007) and (2) a Cauchy-Lognormal-GPD dynamic mixture (Frigessi et al., 2002). In both cases, the static mixture provided a good fit, with Kolmogorov-Smirnov (KS) and Anderson-Darling (AD) tests failing to reject the model.
  • Generalized Beta (GB2) Data: When fitted to highly skewed GB2 data (where all models are mis-specified), the static mixture outperformed the composite and dynamic models in terms of Goodness-of-Fit (GoF) tests and Bayesian Information Criterion (BIC).
  • Computational Cost: The EM-based approach is slower than direct MLE for the composite Lognormal-Pareto model but significantly faster (orders of magnitude) than MLE for the dynamic mixture, which requires numerical integration for the normalizing constant.

• Real-Data Applications:

  • Automobile Claims: The static mixture provided a superior fit compared to pure Lognormal or pure GPD models. It yielded Value-at-Risk (VaR) estimates consistent with empirical quantiles, with lower variability than the dynamic mixture approach.
  • Operational Risk: Similarly, the mixture model fit the heavy-tailed operational loss data better than single-component models. While estimates had wider confidence intervals due to a smaller sample size, the model successfully captured the heavy tail ($\hat{\xi} > 0$) and provided VaR estimates closer to empirical quantiles than competing models.

Significance and Claims
The authors claim that the proposed static Lognormal-GPD mixture offers a practical balance between model flexibility and estimation feasibility. Its primary significance lies in being a fully unsupervised approach that eliminates the need for threshold selection while avoiding the computational bottlenecks of dynamic mixtures.

The paper modestly asserts that the approach is "equivalent in terms of goodness-of-fit" to existing complex distributions but "easier to estimate." It highlights that the model is particularly useful for risk measurement (VaR) and suggests that the posterior probabilities generated by the EM algorithm could be utilized for classification purposes (identifying which observations belong to the tail). The authors acknowledge that while the model is designed for loss data, it may be applicable to other skewed, heavy-tailed phenomena, though they do not claim it is a universal solution for all data types.