Study guide
4.4 Gamma & Inverse Gaussian Models for Claim Severity
CAS Property & Casualty Predictive Analytics PCPA
Track your progress as you read
Finished sections get a check mark.
1 Chapter 1: PCPA Exam & Project Orientation
1.1 PCPA Exam Format, Scoring & Registration 1.2 The PCPA Project: Windows, Deliverables & Grading
2 Chapter 2: Dealing with Data — Sourcing, Importing & Assessing
2.1 Requirements Gathering & Assessing Data Relevance 2.2 Importing, Joining & Reshaping Data Sets 2.3 Exploratory Data Analysis & Evaluating a Data Set 2.4 Data Quality Under ASOP 23 & Modeling Under ASOP 56
3 Chapter 3: Dealing with Data — Conditioning P&C Data for a GLM
3.1 Exposure Bases, Weights & Data Grain 3.2 Missing Data Mechanisms & Imputation Strategies 3.3 Outlier Diagnostics, Trimming & Winsorization 3.4 Variable Transformation & Scaling
4 Chapter 4: Model Diagnostics & Selection — Building a GLM
4.1 Exponential Dispersion Family & Link Functions 4.2 Poisson Regression & Exposure Offsets 4.3 Overdispersion, Quasi-Poisson & Negative Binomial 4.4 Gamma & Inverse Gaussian Models for Claim Severity
5 Chapter 5: Model Diagnostics & Selection — Model Form & Predictors
5.1 Tweedie Compound Poisson-Gamma Modeling 5.2 Binomial & Logistic GLMs for Binary Outcomes 5.3 Categorical Predictors, Base Levels & High-Cardinality Grouping 5.4 Interaction Terms & Non-Linear Effects 5.5 Interpreting Coefficients & Multiplicative Relativities
6 Chapter 6: Model Diagnostics & Selection — Evaluating and Improving the Fit
6.1 Deviance, Likelihood Ratio Tests & Residual Diagnostics 6.2 Stepwise Selection & Information Criteria Penalties 6.3 Multicollinearity, Correlated Predictors & Aliasing 6.4 Under-fitting, Over-fitting & Cross-Validation
7 Chapter 7: Model Diagnostics & Selection — Validation & Model Choice
7.1 Actual vs. Predicted, Calibration & Loss Ratio Charts 7.2 Simple Quantile Plots & Measuring Lift 7.3 Double Lift Charts & Competing Model Evaluation 7.4 Gini Index, Lorenz Curves & ROC Curves
8 Chapter 8: Model Interpretation & Presentation — Tables, Charts & Exhibits
8.1 Designing Tables & Statistical Exhibits 8.2 Chart Selection & Graphical Perception 8.3 How Charts Mislead 8.4 One-Way, Two-Way & Partial Dependence Exhibits
9 Chapter 9: Model Interpretation & Presentation — Communicating the Analysis
9.1 Communicating Methodology & Modeling Decisions 9.2 Explaining Why a Variable Was or Was Not Included 9.3 Communicating to Non-Technical Audiences 9.4 The PCPA Technical Report: Structure & Rubric
4.4 Gamma & Inverse Gaussian Models for Claim Severity
Key Takeaways Explain This Topic AI
Claim severity models the conditional loss size given that a claim has occurred (Y > 0); policies with zero claims must be excluded from severity estimation datasets.
When modeling average claim severity at the policy or rating cell level (Y_bar = Total Loss / Claim Count), observations must be weighted by prior weights equal to the claim count omega_i = N_i.
The Gamma distribution exhibits a variance function V(mu) = mu^2, leading to a constant coefficient of variation CV = sqrt(phi) across all rating risk classes regardless of expected loss magnitude.
The Inverse Gaussian distribution features a steeper variance function V(mu) = mu^3 and non-constant CV = sqrt(phi * mu), creating a substantially heavier right tail that better captures commercial liability shock losses.
The log link g(mu) = ln(mu) is universally favored over canonical links for severity because it strictly guarantees positive loss predictions and directly produces multiplicative actuarial rating relativities.
Last updated: September 2026
Severity Modeling Principles in Actuarial Practice
In property and casualty (P&C) ratemaking, pure premium (loss cost) represents the expected loss per unit of exposure. Under the classical two-stage frequency-severity framework, pure premium decomposes multiplicatively into the expected claim frequency and the expected conditional claim severity: Pure Premium = E [ Loss Exposure ] = E [ Claims Exposure ] × E [ Loss Claims | Claims > 0 ] = Frequency × Severity \text{Pure Premium} = \mathbb{E}\left[\frac{\text{Loss}}{\text{Exposure}}\right] = \mathbb{E}\left[\frac{\text{Claims}}{\text{Exposure}}\right] \times \mathbb{E}\left[\frac{\text{Loss}}{\text{Claims}} \;\middle|\; \text{Claims} > 0\right] = \text{Frequency} \times \text{Severity} Pure Premium = E [ Exposure Loss ] = E [ Exposure Claims ] × E [ Claims Loss Claims > 0 ] = Frequency × Severity
While frequency models analyze discrete claim counts across all exposed policies (including zero-claim policies), claim severity models operate strictly on the conditional distribution of claim sizes given that a claim has occurred ( Y i > 0 Y_i > 0 Y i > 0 ). ALL EXPOSED POLICIES
- │
- ┌──────────────────────┴──────────────────────┐
- ▼ ▼
- No Claims (Y = 0) Claim Occurred (Y > 0)
- (Used in Frequency Model) (Used in Severity Model)
- │
- ┌──────────────┴──────────────┐
- ▼ ▼
- Claim-Level Policy-Level
- Record Grain Average Severity
- (1 row per claim) (Y_bar = Loss / N, Weight = N)
- Data Grain and Actuarial Weighting
Actuaries fit severity models at one of two distinct data grains:
Individual Claim Grain (Transaction Level) : Each observation represents a single closed or settled claim ( y i y_i y i ). The model assigns equal prior weights ( ω i = 1 \omega_i = 1 ω i = 1 ) unless specific inflation trend adjustments or credibility weights apply.
Policy or Risk-Cell Grain (Aggregated Level) : Each observation represents a policy or rating cell that incurred one or more claims ( N i ≥ 1 N_i \ge 1 N i ≥ 1 ). The target response is the average claim severity : Y ˉ i = S i N i = ∑ j = 1 N i Z i j N i \bar{Y}_i = \frac{S_i}{N_i} = \frac{\sum_{j=1}^{N_i} Z_{ij}}{N_i} Y ˉ i = N i S i = N i ∑ j = 1 N i Z ij
Where S i S_i S i is the aggregate loss and Z i j Z_{ij} Z ij is the amount of the j j j -th individual claim on policy i i i .
If individual claims Z i j Z_{ij} Z ij are independent and identically distributed with mean μ i \mu_i μ i and variance Var ( Z i j ) = ϕ V ( μ i ) \text{Var}(Z_{ij}) = \phi V(\mu_i) Var ( Z ij ) = ϕ V ( μ i ) , the variance of the average severity is: Var ( Y ˉ i ) = Var ( 1 N i ∑ j = 1 N i Z i j ) = 1 N i 2 ∑ j = 1 N i Var ( Z i j ) = N i ϕ V ( μ i ) N i 2 = ϕ N i V ( μ i ) \text{Var}(\bar{Y}_i) = \text{Var}\left( \frac{1}{N_i} \sum_{j=1}^{N_i} Z_{ij} \right) = \frac{1}{N_i^2} \sum_{j=1}^{N_i} \text{Var}(Z_{ij}) = \frac{N_i \phi V(\mu_i)}{N_i^2} = \frac{\phi}{N_i} V(\mu_i) Var ( Y ˉ i ) = Var ( N i 1 j = 1 ∑ N i Z ij ) = N i 2 1 j = 1 ∑ N i Var ( Z ij ) = N i 2 N i ϕ V ( μ i ) = N i ϕ V ( μ i )
In the Exponential Dispersion Family (EDF), the dispersion function is parameterized as a ( ϕ ) = ϕ ω i a(\phi) = \frac{\phi}{\omega_i} a ( ϕ ) = ω i ϕ . Comparing this with the variance equation demonstrates that the prior weight ω i \omega_i ω i must strictly equal the claim count N i N_i N i : ω i = N i \omega_i = N_i ω i = N i
An observation with 10 10 10 claims provides ten times the information and exhibits one-tenth the sampling variance of an observation with 1 1 1 claim. Weighting by claim count ensures that parameter estimation is unbiased and statistically efficient.
The Gamma Distribution for Claim Severity
The Gamma distribution is the standard workhorse for modeling P&C insurance claim severity. Its support is strictly positive ( y > 0 y > 0 y > 0 ), and its right-skewed profile naturally reflects the asymmetry of property damage and collision losses.
Probability Density Function and EDF Parameterization
In classical statistics, the Gamma distribution is parameterized by shape parameter α > 0 \alpha > 0 α > 0 and scale parameter θ > 0 \theta > 0 θ > 0 . In Generalized Linear Models, actuaries reparameterize the distribution in terms of its mean μ = E [ Y ] = α θ \mu = \mathbb{E}[Y] = \alpha \theta μ = E [ Y ] = α θ and dispersion parameter ϕ = 1 α > 0 \phi = \frac{1}{\alpha} > 0 ϕ = α 1 > 0 : f ( y ; μ , ϕ ) = 1 Γ ( 1 / ϕ ) ( 1 ϕ μ ) 1 / ϕ y 1 / ϕ − 1 exp ( − y ϕ μ ) , y > 0 f(y; \mu, \phi) = \frac{1}{\Gamma(1/\phi)} \left( \frac{1}{\phi \mu} \right)^{1/\phi} y^{1/\phi - 1} \exp\left( -\frac{y}{\phi \mu} \right), \quad y > 0 f ( y ; μ , ϕ ) = Γ ( 1/ ϕ ) 1 ( ϕ μ 1 ) 1/ ϕ y 1/ ϕ − 1 exp ( − ϕ μ y ) , y > 0
To express this in canonical Exponential Dispersion Family form: f ( y ; θ c , ϕ ) = exp ( y θ c − b ( θ c ) ϕ + c ( y , ϕ ) ) f(y; \theta_c, \phi) = \exp\left( \frac{y \theta_c - b(\theta_c)}{\phi} + c(y, \phi) \right) f ( y ; θ c , ϕ ) = exp ( ϕ y θ c − b ( θ c ) + c ( y , ϕ ) )
Where:
Natural (canonical) parameter: θ c = − 1 μ ⟹ μ = − 1 θ c \theta_c = -\frac{1}{\mu} \implies \mu = -\frac{1}{\theta_c} θ c = − μ 1 ⟹ μ = − θ c 1
Cumulant function: b ( θ c ) = − ln ( − θ c ) = ln ( μ ) b(\theta_c) = -\ln(-\theta_c) = \ln(\mu) b ( θ c ) = − ln ( − θ c ) = ln ( μ )
Dispersion parameter: a ( ϕ ) = ϕ ω i a(\phi) = \frac{\phi}{\omega_i} a ( ϕ ) = ω i ϕ
Normalizing function: c ( y , ϕ ) = 1 − ϕ ϕ ln ( y ) − ln ( ϕ ) ϕ − ln Γ ( 1 ϕ ) c(y, \phi) = \frac{1 - \phi}{\phi} \ln(y) - \frac{\ln(\phi)}{\phi} - \ln\Gamma\left(\frac{1}{\phi}\right) c ( y , ϕ ) = ϕ 1 − ϕ ln ( y ) − ϕ l n ( ϕ ) − ln Γ ( ϕ 1 )
Derivation of the Variance Function
The mean and variance functions derive directly from the cumulant function b ( θ c ) b(\theta_c) b ( θ c ) : μ = b ′ ( θ c ) = − 1 − θ c ( − 1 ) = − 1 θ c \mu = b'(\theta_c) = -\frac{1}{-\theta_c}(-1) = -\frac{1}{\theta_c} μ = b ′ ( θ c ) = − − θ c 1 ( − 1 ) = − θ c 1 V ( μ ) = b ′ ′ ( θ c ) = 1 θ c 2 = ( − 1 μ ) − 2 = μ 2 V(\mu) = b''(\theta_c) = \frac{1}{\theta_c^2} = \left( -\frac{1}{\mu} \right)^{-2} = \mu^2 V ( μ ) = b ′′ ( θ c ) = θ c 2 1 = ( − μ 1 ) − 2 = μ 2
Thus, the variance of a Gamma random variable with prior weight ω i \omega_i ω i is: Var ( Y i ) = ϕ ω i V ( μ i ) = ϕ ω i μ i 2 \text{Var}(Y_i) = \frac{\phi}{\omega_i} V(\mu_i) = \frac{\phi}{\omega_i} \mu_i^2 Var ( Y i ) = ω i ϕ V ( μ i ) = ω i ϕ μ i 2
The Constant Coefficient of Variation (CV) Property
A critical mathematical property of the Gamma GLM is that its Coefficient of Variation ( C V CV C V ) is strictly constant across all risk profiles : C V = σ μ = Var ( Y ) μ = ϕ μ 2 μ = ϕ = 1 α CV = \frac{\sigma}{\mu} = \frac{\sqrt{\text{Var}(Y)}}{\mu} = \frac{\sqrt{\phi \mu^2}}{\mu} = \sqrt{\phi} = \frac{1}{\sqrt{\alpha}} C V = μ σ = μ Var ( Y ) = μ ϕ μ 2 = ϕ = α 1
This constant C V CV C V property has profound actuarial meaning:
Proportional uncertainty is identical across rating tiers: A risk group with an expected claim severity of $1,000 (e.g., minor auto bumper repairs) has a standard deviation of ϕ × 1 , 000 \sqrt{\phi} \times 1,000 ϕ × 1 , 000 . A risk group with an expected severity of $20,000 (e.g., luxury sports car repairs) has a standard deviation of ϕ × 20 , 000 \sqrt{\phi} \times 20,000 ϕ × 20 , 000 .
If ϕ = 0.25 \phi = 0.25 ϕ = 0.25 , then C V = 0.25 = 0.50 CV = \sqrt{0.25} = 0.50 C V = 0.25 = 0.50 ( 50 % 50\% 50% ) for both risk classes.
This reflects property lines where loss volatility scales quadratically with the value of the underlying asset.
Link Functions for Severity: Log Link vs. Canonical Link
The canonical link function for the Gamma distribution equates the linear predictor η i = x i T β \eta_i = \mathbf{x}_i^T \boldsymbol{\beta} η i = x i T β to the natural parameter θ c = − 1 / μ i \theta_c = -1/\mu_i θ c = − 1/ μ i : g ( μ i ) = − 1 μ i ( or operational inverse link g ( μ i ) = 1 μ i ) g(\mu_i) = -\frac{1}{\mu_i} \quad \left(\text{or operational inverse link } g(\mu_i) = \frac{1}{\mu_i}\right) g ( μ i ) = − μ i 1 ( or operational inverse link g ( μ i ) = μ i 1 )
Why Actuaries Universally Reject the Canonical Link for Severity
Despite its mathematical appeal (such as minimal sufficient statistics and Newton-Raphson equivalence), the canonical reciprocal link is virtually never used in commercial insurance pricing:
Risk of Negative or Undefined Fitted Values : Under the inverse link, μ ^ i = 1 x i T β ^ \hat{\mu}_i = \frac{1}{\mathbf{x}_i^T \hat{\boldsymbol{\beta}}} μ ^ i = x i T β ^ 1 . If the linear predictor evaluates to zero or a negative number for any policyholder profile, the model predicts an infinite or negative expected claim severity. In insurance, negative claim severities are a physical impossibility that will crash automated rating engines.
Non-Multiplicative Rating Structure : The inverse link produces an additive model on the reciprocal scale, which cannot be incorporated into standard multiplicative insurance rating manuals.
The Operational Dominance of the Log Link
Actuaries universally pair the Gamma distribution with the log link function : g ( μ i ) = ln ( μ i ) = x i T β ⟺ μ i = exp ( x i T β ) g(\mu_i) = \ln(\mu_i) = \mathbf{x}_i^T \boldsymbol{\beta} \iff \mu_i = \exp(\mathbf{x}_i^T \boldsymbol{\beta}) g ( μ i ) = ln ( μ i ) = x i T β ⟺ μ i = exp ( x i T β )
This specification provides two indispensable benefits:
Strict Positivity : Because exp ( η ) > 0 \exp(\eta) > 0 exp ( η ) > 0 for all η ∈ ( − ∞ , ∞ ) \eta \in (-\infty, \infty) η ∈ ( − ∞ , ∞ ) , predicted severity is mathematically guaranteed to be strictly positive regardless of predictor combinations.
Multiplicative Relativities : The model decomposes directly into standard actuarial rating relativities: μ i = exp ( β 0 ) × exp ( β 1 x i 1 ) × exp ( β 2 x i 2 ) × ⋯ × exp ( β k x i k ) \mu_i = \exp(\beta_0) \times \exp(\beta_1 x_{i1}) \times \exp(\beta_2 x_{i2}) \times \dots \times \exp(\beta_k x_{ik}) μ i = exp ( β 0 ) × exp ( β 1 x i 1 ) × exp ( β 2 x i 2 ) × ⋯ × exp ( β k x ik )
The base severity is Base Rate = exp ( β 0 ) \text{Base Rate} = \exp(\beta_0) Base Rate = exp ( β 0 ) .
Each categorical attribute contributes a multiplicative relativity Relativity j = exp ( β j ) \text{Relativity}_j = \exp(\beta_j) Relativity j = exp ( β j ) .
The Inverse Gaussian Distribution for Heavy-Tailed Severity
While the Gamma distribution fits well for well-behaved property damage claims, it frequently underestimates loss costs in commercial lines characterized by severe shock losses, long-tail casualty litigation, and catastrophic personal injury. For these lines, the Inverse Gaussian (IG) distribution provides a superior alternative.
Probability Density Function and EDF Parameterization
The Inverse Gaussian distribution (also known as the Wald distribution) has probability density function: f ( y ; μ , ϕ ) = 1 2 π ϕ y 3 exp ( − ( y − μ ) 2 2 ϕ μ 2 y ) , y > 0 f(y; \mu, \phi) = \frac{1}{\sqrt{2\pi \phi y^3}} \exp\left( -\frac{(y - \mu)^2}{2 \phi \mu^2 y} \right), \quad y > 0 f ( y ; μ , ϕ ) = 2 π ϕ y 3 1 exp ( − 2 ϕ μ 2 y ( y − μ ) 2 ) , y > 0
In Exponential Dispersion Family notation: f ( y ; θ c , ϕ ) = exp ( y θ c − b ( θ c ) ϕ + c ( y , ϕ ) ) f(y; \theta_c, \phi) = \exp\left( \frac{y \theta_c - b(\theta_c)}{\phi} + c(y, \phi) \right) f ( y ; θ c , ϕ ) = exp ( ϕ y θ c − b ( θ c ) + c ( y , ϕ ) )
Where:
Natural parameter: θ c = − 1 2 μ 2 ⟹ μ = ( − 2 θ c ) − 1 / 2 \theta_c = -\frac{1}{2\mu^2} \implies \mu = (-2\theta_c)^{-1/2} θ c = − 2 μ 2 1 ⟹ μ = ( − 2 θ c ) − 1/2
Cumulant function: b ( θ c ) = − − 2 θ c = − 1 μ b(\theta_c) = -\sqrt{-2\theta_c} = -\frac{1}{\mu} b ( θ c ) = − − 2 θ c = − μ 1
Dispersion parameter: a ( ϕ ) = ϕ ω i a(\phi) = \frac{\phi}{\omega_i} a ( ϕ ) = ω i ϕ
Normalizing function: c ( y , ϕ ) = − 1 2 ln ( 2 π ϕ y 3 ) − 1 2 ϕ y c(y, \phi) = -\frac{1}{2}\ln(2\pi \phi y^3) - \frac{1}{2\phi y} c ( y , ϕ ) = − 2 1 ln ( 2 π ϕ y 3 ) − 2 ϕ y 1
Derivation of the Variance Function
Differentiating the cumulant function yields the mean and variance: μ = b ′ ( θ c ) = − 1 2 ( − 2 θ c ) − 1 / 2 ( − 2 ) = ( − 2 θ c ) − 1 / 2 = μ \mu = b'(\theta_c) = -\frac{1}{2}(-2\theta_c)^{-1/2}(-2) = (-2\theta_c)^{-1/2} = \mu μ = b ′ ( θ c ) = − 2 1 ( − 2 θ c ) − 1/2 ( − 2 ) = ( − 2 θ c ) − 1/2 = μ V ( μ ) = b ′ ′ ( θ c ) = − 1 2 ( − 2 θ c ) − 3 / 2 ( − 2 ) = ( − 2 θ c ) − 3 / 2 = ( 1 μ 2 ) − 3 / 2 = μ 3 V(\mu) = b''(\theta_c) = -\frac{1}{2}(-2\theta_c)^{-3/2}(-2) = (-2\theta_c)^{-3/2} = \left(\frac{1}{\mu^2}\right)^{-3/2} = \mu^3 V ( μ ) = b ′′ ( θ c ) = − 2 1 ( − 2 θ c ) − 3/2 ( − 2 ) = ( − 2 θ c ) − 3/2 = ( μ 2 1 ) − 3/2 = μ 3
Thus, the variance of an Inverse Gaussian random variable is cubic in the mean: Var ( Y i ) = ϕ ω i V ( μ i ) = ϕ ω i μ i 3 \text{Var}(Y_i) = \frac{\phi}{\omega_i} V(\mu_i) = \frac{\phi}{\omega_i} \mu_i^3 Var ( Y i ) = ω i ϕ V ( μ i ) = ω i ϕ μ i 3
Non-Constant Coefficient of Variation in Inverse Gaussian
Because the variance function is V ( μ ) = μ 3 V(\mu) = \mu^3 V ( μ ) = μ 3 , the Coefficient of Variation is not constant ; it increases with the square root of the mean: C V = σ μ = ϕ μ 3 μ = ϕ μ CV = \frac{\sigma}{\mu} = \frac{\sqrt{\phi \mu^3}}{\mu} = \sqrt{\phi \mu} C V = μ σ = μ ϕ μ 3 = ϕ μ
This behavior matches long-tail liability insurance: high-severity risk classes (such as heavy commercial trucking or hazardous industrial manufacturing) exhibit not only higher absolute dollar variances, but also proportionately greater relative volatility than low-severity risk classes.
Tail Behavior Comparison: Gamma vs. Inverse Gaussian
The fundamental actuarial distinction between the Gamma and Inverse Gaussian distributions lies in their asymptotic tail decay. Consider the right tail behavior as claim size y → ∞ y \to \infty y → ∞ : Gamma Density Decay: f Gamma ( y ) ∝ y 1 / ϕ − 1 exp ( − y ϕ μ ) \text{Gamma Density Decay:} \quad f_{\text{Gamma}}(y) \propto y^{1/\phi - 1} \exp\left(-\frac{y}{\phi \mu}\right) Gamma Density Decay: f Gamma ( y ) ∝ y 1/ ϕ − 1 exp ( − ϕ μ y ) Inverse Gaussian Density Decay: f IG ( y ) ∝ y − 3 / 2 exp ( − y 2 ϕ μ 2 ) \text{Inverse Gaussian Density Decay:} \quad f_{\text{IG}}(y) \propto y^{-3/2} \exp\left(-\frac{y}{2\phi \mu^2}\right) Inverse Gaussian Density Decay: f IG ( y ) ∝ y − 3/2 exp ( − 2 ϕ μ 2 y ) Density f(y)
- │
- │ ▲
- │ ╱ ╲
- │ ╱ ╲ Gamma (Decays faster in right tail: V(mu) = mu^2)
- │ ╱ ╲─────────────────────────┐
- │ ╱ └───┐
- │ ╱ Inverse Gaussian └──────────────────
- │╱ (Heavy right tail: V(mu) = mu^3) ┌──────────────
- └────────────────────────────────────────────┴───────────────► Claim Size (y)
- Why Inverse Gaussian Has a Heavier Right Tail
Variance Growth Rate : Because V ( μ ) = μ 3 V(\mu) = \mu^3 V ( μ ) = μ 3 for Inverse Gaussian versus V ( μ ) = μ 2 V(\mu) = \mu^2 V ( μ ) = μ 2 for Gamma, the conditional variance grows at a much faster rate as expected losses increase.
Shock Loss Sensitivity : The Inverse Gaussian distribution allocates substantially higher probability mass to extreme positive values ( y ≫ μ y \gg \mu y ≫ μ ).
Actuarial Suitability by Line of Business :
Line of Business Typical Claim Profile Recommended Severity Model Actuarial Rationale
Personal Auto Collision Capped by vehicle actual cash value (ACV); well-behaved, moderate skewness Gamma Loss amounts have natural physical ceilings; constant C V CV C V is realistic.
Personal Homeowners (Property) Building structure limits cap structural losses; moderate tail Gamma Most claims are localized repairs; extreme fire totals are handled via policy limits.
Commercial Auto Bodily Injury High policy limits ($1M+), severe medical litigation, life-care plans Inverse Gaussian Extreme right tail with frequent shock verdicts; cubic variance captures high-tier risk volatility.
Commercial General Liability Long latency, toxic torts, multi-party construction defect litigation Inverse Gaussian High proportion of multi-million dollar shock claims; Gamma severely under-predicts right tail.
Medical Malpractice High severity, low frequency, catastrophic lifetime medical expenses Inverse Gaussian Skewness is extreme; Inverse Gaussian prevents underpricing high-risk surgical specialties.
Model Selection and Actuarial Evaluation Framework
When evaluating competing severity distributions (Gamma vs. Inverse Gaussian), actuaries apply a multi-faceted statistical and operational framework:
1. Unit Deviance Comparison
The unscaled unit deviance d ( y i , μ i ) d(y_i, \mu_i) d ( y i , μ i ) reflects the discrepancy between the observed claim size and the fitted mean:
Gamma Unit Deviance : d Gamma ( y , μ ) = 2 [ − ln ( y μ ) + y − μ μ ] d_{\text{Gamma}}(y, \mu) = 2 \left[ -\ln\left(\frac{y}{\mu}\right) + \frac{y - \mu}{\mu} \right] d Gamma ( y , μ ) = 2 [ − ln ( μ y ) + μ y − μ ]
Inverse Gaussian Unit Deviance : d IG ( y , μ ) = ( y − μ ) 2 μ 2 y d_{\text{IG}}(y, \mu) = \frac{(y - \mu)^2}{\mu^2 y} d IG ( y , μ ) = μ 2 y ( y − μ ) 2
The total deviance D ( y , μ ^ ) = ∑ i = 1 n ω i d ( y i , μ ^ i ) D(\mathbf{y}, \hat{\boldsymbol{\mu}}) = \sum_{i=1}^n \omega_i d(y_i, \hat{\mu}_i) D ( y , μ ^ ) = ∑ i = 1 n ω i d ( y i , μ ^ i ) summarizes overall discrepancy.
2. Information Criteria (AIC and BIC)
Because Gamma and Inverse Gaussian models have identical parameter counts for the same predictor set, comparing log-likelihoods directly indicates relative goodness-of-fit: AIC = − 2 ln ( L ) + 2 k \text{AIC} = -2\ln(L) + 2k AIC = − 2 ln ( L ) + 2 k BIC = − 2 ln ( L ) + k ln ( n ) \text{BIC} = -2\ln(L) + k\ln(n) BIC = − 2 ln ( L ) + k ln ( n )
The model achieving lower AIC/BIC is preferred. In severe liability lines, Inverse Gaussian frequently achieves an AIC improvement of hundreds or thousands of points over Gamma due to superior tail fit.
3. Normal Q-Q Plots of Deviance Residuals
Deviance residuals for observation i i i are defined as: r D , i = sign ( y i − μ ^ i ) ω i d ( y i , μ ^ i ) r_{D, i} = \text{sign}(y_i - \hat{\mu}_i) \sqrt{\omega_i d(y_i, \hat{\mu}_i)} r D , i = sign ( y i − μ ^ i ) ω i d ( y i , μ ^ i )
Actuaries plot the ordered deviance residuals against standard normal theoretical quantiles. If the Gamma model's residuals curve sharply upward in the upper right quadrant, the observed data possesses a heavier tail than the Gamma distribution can accommodate. Switching to an Inverse Gaussian model often straightens the upper tail along the 45-degree reference line.
Actuarial Exam Traps & Pitfalls
Warning
Trap 1: Including Zero-Claim Policies in Severity Models Both Gamma and Inverse Gaussian distributions have strictly positive support ( y > 0 y > 0 y > 0 ). If an actuary mistakenly includes policies with zero claims ( y i = 0 y_i = 0 y i = 0 ), evaluating the density or deviance requires computing ln ( 0 ) \ln(0) ln ( 0 ) or division by zero, causing the estimation routine to crash. Severity models must strictly filter data to y i > 0 y_i > 0 y i > 0 .
Warning
Trap 2: Omitting Prior Weights on Aggregated Severity Data When modeling average claim severity at the rating cell level ( Y ˉ i = S i / N i \bar{Y}_i = S_i / N_i Y ˉ i = S i / N i ), failure to set prior weights weights = N treats a cell with 1,000 claims as having the same sampling reliability as a cell with 1 claim. This inflates the parameter estimation variance and distorts rating relativities.
Warning
Trap 3: Assuming Inverse Gaussian Has a Constant Coefficient of Variation On exam questions, candidates frequently assume all severity GLMs have constant C V CV C V . Remember: only the Gamma distribution has constant C V = ϕ CV = \sqrt{\phi} C V = ϕ . The Inverse Gaussian distribution has C V = ϕ μ CV = \sqrt{\phi \mu} C V = ϕ μ , which scales directly with the square root of the expected claim size.
Warning
Trap 4: Using the Canonical Inverse Link for Gamma Severity Never specify the canonical link g ( μ ) = − 1 / μ g(\mu) = -1/\mu g ( μ ) = − 1/ μ (or 1 / μ 1/\mu 1/ μ ) for ratemaking. If the linear predictor evaluates to ≤ 0 \le 0 ≤ 0 for favorable rating tiers, it produces negative or infinite expected severities. Always use the log link g ( μ ) = ln ( μ ) g(\mu) = \ln(\mu) g ( μ ) = ln ( μ ) .