Study guide
5.1 Tweedie Compound Poisson-Gamma Modeling
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
5.1 Tweedie Compound Poisson-Gamma Modeling
Key Takeaways Explain This Topic AI
The Tweedie Compound Poisson-Gamma distribution models aggregate loss Y = sum_{i=1}^N Z_i with Poisson claim frequency N ~ Poisson(lambda) and Gamma severity Z_i ~ Gamma(alpha, theta).
The Tweedie variance function is V(mu) = mu^p where the power parameter strictly satisfies 1 < p < 2, creating an exact point mass at zero P(Y = 0) = exp(-lambda) > 0 for non-claim policies.
The Tweedie parameters (mu, phi, p) map bijectively to Poisson-Gamma parameters: shape alpha = (2 - p)/(p - 1), Poisson mean lambda = mu^(2 - p) / [phi * (2 - p)], and Gamma scale theta = phi * (p - 1) * mu^(p - 1).
Direct Tweedie pure premium modeling directly optimizes loss cost deviance in a single model, eliminating the operational overhead of managing separate frequency and severity pipelines.
The index parameter p cannot be solved via standard Iteratively Reweighted Least Squares (IRLS) and must be estimated through profile log-likelihood grid search across the interval (1.01, 1.99).
Last updated: September 2026
The Compound Poisson-Gamma Process in P&C Insurance
In property and casualty insurance, pure premium (or loss cost) represents aggregate losses per unit of earned exposure. Traditional ratemaking structures pure premium via two separate Generalized Linear Models: a Poisson model for claim frequency and a Gamma model for claim severity. However, casualty actuaries often model pure premium directly in a single step using the Tweedie Compound Poisson-Gamma distribution .
Consider an insurance portfolio where an insured unit over a defined coverage period experiences an unknown number of claims N N N , each generating a loss amount Z i Z_i Z i : Y = ∑ i = 1 N Z i Y = \sum_{i=1}^N Z_i Y = i = 1 ∑ N Z i
Where:
Claim frequency N N N follows a Poisson distribution with intensity parameter λ > 0 \lambda > 0 λ > 0 : P ( N = n ) = e − λ λ n n ! , n ∈ { 0 , 1 , 2 , … } P(N = n) = \frac{e^{-\lambda} \lambda^n}{n!}, \quad n \in \{0, 1, 2, \dots\} P ( N = n ) = n ! e − λ λ n , n ∈ { 0 , 1 , 2 , … }
Individual claim severities Z 1 , Z 2 , … , Z N Z_1, Z_2, \dots, Z_N Z 1 , Z 2 , … , Z N are independent, identically distributed Gamma random variables with shape parameter α > 0 \alpha > 0 α > 0 and scale parameter θ > 0 \theta > 0 θ > 0 : Z i ∼ i i d Gamma ( α , θ ) , E [ Z i ] = α θ , Var ( Z i ) = α θ 2 Z_i \stackrel{iid}{\sim} \text{Gamma}(\alpha, \theta), \quad \mathbb{E}[Z_i] = \alpha\theta, \quad \text{Var}(Z_i) = \alpha\theta^2 Z i ∼ ii d Gamma ( α , θ ) , E [ Z i ] = α θ , Var ( Z i ) = α θ 2
The count N N N and the individual claim amounts Z i Z_i Z i are mutually independent.
If N = 0 N = 0 N = 0 , the aggregate loss is identically zero: Y = 0 Y = 0 Y = 0 .
Mean and Variance of Aggregate Losses
Applying the Law of Total Expectation: μ = E [ Y ] = E [ E [ Y ∣ N ] ] = E [ N ⋅ E [ Z ] ] = E [ N ] ⋅ E [ Z ] = λ α θ \mu = \mathbb{E}[Y] = \mathbb{E}\Big[ \mathbb{E}[Y \mid N] \Big] = \mathbb{E}[N \cdot \mathbb{E}[Z]] = \mathbb{E}[N] \cdot \mathbb{E}[Z] = \lambda \alpha \theta μ = E [ Y ] = E [ E [ Y ∣ N ] ] = E [ N ⋅ E [ Z ]] = E [ N ] ⋅ E [ Z ] = λ α θ
Applying the Law of Total Variance (Eve's Law): Var ( Y ) = E [ Var ( Y ∣ N ) ] + Var ( E [ Y ∣ N ] ) \text{Var}(Y) = \mathbb{E}\Big[ \text{Var}(Y \mid N) \Big] + \text{Var}\Big( \mathbb{E}[Y \mid N] \Big) Var ( Y ) = E [ Var ( Y ∣ N ) ] + Var ( E [ Y ∣ N ] ) Var ( Y ) = E [ N ⋅ Var ( Z ) ] + Var ( N ⋅ E [ Z ] ) = E [ N ] Var ( Z ) + ( E [ Z ] ) 2 Var ( N ) \text{Var}(Y) = \mathbb{E}[N \cdot \text{Var}(Z)] + \text{Var}(N \cdot \mathbb{E}[Z]) = \mathbb{E}[N] \text{Var}(Z) + (\mathbb{E}[Z])^2 \text{Var}(N) Var ( Y ) = E [ N ⋅ Var ( Z )] + Var ( N ⋅ E [ Z ]) = E [ N ] Var ( Z ) + ( E [ Z ] ) 2 Var ( N )
Substituting the Poisson and Gamma moments: Var ( Y ) = λ ( α θ 2 ) + ( α θ ) 2 ( λ ) = λ α θ 2 ( 1 + α ) \text{Var}(Y) = \lambda (\alpha \theta^2) + (\alpha \theta)^2 (\lambda) = \lambda \alpha \theta^2 (1 + \alpha) Var ( Y ) = λ ( α θ 2 ) + ( α θ ) 2 ( λ ) = λ α θ 2 ( 1 + α )
The Tweedie Exponential Dispersion Family Representation
In 1984, British statistician Maurice Tweedie demonstrated that this Compound Poisson-Gamma process belongs to the Exponential Dispersion Family (EDF) with a specific power variance function: V ( μ ) = μ p , where 1 < p < 2 V(\mu) = \mu^p, \quad \text{where } 1 < p < 2 V ( μ ) = μ p , where 1 < p < 2 Var ( Y ) = ϕ V ( μ ) = ϕ μ p \text{Var}(Y) = \phi V(\mu) = \phi \mu^p Var ( Y ) = ϕ V ( μ ) = ϕ μ p
Where ϕ \phi ϕ is the dispersion parameter and p p p is known as the Tweedie index (power) parameter .
Spectrum of Tweedie Distributions
The Tweedie family encompasses several classical distributions depending on the value of p p p :
Tweedie Power ( p p p ) Underlying Distribution Support Insurance Pricing Application
p = 0 p = 0 p = 0 Normal (Gaussian) ( − ∞ , ∞ ) (-\infty, \infty) ( − ∞ , ∞ ) Benchmark linear regression (rarely used for losses)
p = 1 p = 1 p = 1 Poisson { 0 , 1 , 2 , … } \{0, 1, 2, \dots\} { 0 , 1 , 2 , … } Claim frequency modeling (count data)
1 < p < 2 1 < p < 2 1 < p < 2 Compound Poisson-Gamma [ 0 , ∞ ) [0, \infty) [ 0 , ∞ ) Pure premium / loss cost modeling (mass at zero)
- p = 2 p = 2 p = 2 Gamma ( 0 , ∞ ) (0, \infty) ( 0 , ∞ ) Claim severity modeling (strictly positive)
- p = 3 p = 3 p = 3 Inverse Gaussian ( 0 , ∞ ) (0, \infty) ( 0 , ∞ ) Heavy-tailed claim severity modeling
p > 2 p > 2 p > 2 ( p ≠ 3 p \ne 3 p = 3 ) Positive Stable ( 0 , ∞ ) (0, \infty) ( 0 , ∞ ) Extreme value and catastrophe loss modeling TWEEDIE INDEX SPECTRUM (p)
- p = 0 p = 1 1 < p < 2 p = 2 p = 3
- ───────┼─────────────────┼───────────────────┼──────────────────┼───────────────┼───────►
- Normal Poisson TWEEDIE Gamma Inverse
- (Linear) (Frequency) (Pure Premium) (Severity) Gaussian
- Compound P-G (Heavy Tail)
- [Mass at Zero!]
- The Point Mass at Zero
The defining feature of the Tweedie distribution with 1 < p < 2 1 < p < 2 1 < p < 2 is that it is not purely continuous , but rather a mixed discrete-continuous distribution . It exhibits an exact point mass at zero corresponding to policies that incurred no claims ( N = 0 N = 0 N = 0 ): P ( Y = 0 ) = P ( N = 0 ) = e − λ > 0 P(Y = 0) = P(N = 0) = e^{-\lambda} > 0 P ( Y = 0 ) = P ( N = 0 ) = e − λ > 0
For y > 0 y > 0 y > 0 , the distribution is strictly continuous with a smooth, right-skewed density. In a personal auto portfolio where 90 % 90\% 90% of drivers have no claims during the year, the empirical pure premium distribution has a spike of probability at $0 ($90%$ mass) followed by a continuous positive distribution for the 10 % 10\% 10% who incurred losses. The Tweedie distribution matches this physical reality.
Mathematical Mapping: Tweedie vs. Poisson-Gamma Parameters
There is an exact, invertible algebraic mapping between the Tweedie GLM parameters ( μ , ϕ , p ) (\mu, \phi, p) ( μ , ϕ , p ) and the underlying compound Poisson-Gamma parameters ( λ , α , θ ) (\lambda, \alpha, \theta) ( λ , α , θ ) :
Derivation of the Transformation
We equate the Tweedie mean and variance to the Poisson-Gamma mean and variance: μ = λ α θ \mu = \lambda \alpha \theta μ = λ α θ Var ( Y ) = ϕ μ p = λ α θ 2 ( 1 + α ) \text{Var}(Y) = \phi \mu^p = \lambda \alpha \theta^2 (1 + \alpha) Var ( Y ) = ϕ μ p = λ α θ 2 ( 1 + α )
Dividing the variance by the mean: Var ( Y ) μ = ϕ μ p μ = ϕ μ p − 1 = λ α θ 2 ( 1 + α ) λ α θ = θ ( 1 + α ) \frac{\text{Var}(Y)}{\mu} = \frac{\phi \mu^p}{\mu} = \phi \mu^{p-1} = \frac{\lambda \alpha \theta^2 (1 + \alpha)}{\lambda \alpha \theta} = \theta (1 + \alpha) μ Var ( Y ) = μ ϕ μ p = ϕ μ p − 1 = λ α θ λ α θ 2 ( 1 + α ) = θ ( 1 + α ) θ = ϕ μ p − 1 1 + α \theta = \frac{\phi \mu^{p-1}}{1 + \alpha} θ = 1 + α ϕ μ p − 1
Substituting θ \theta θ back into the mean expression μ = λ α θ \mu = \lambda \alpha \theta μ = λ α θ : μ = λ α ( ϕ μ p − 1 1 + α ) ⟹ λ = μ 2 − p ϕ ( 1 + α α ) \mu = \lambda \alpha \left( \frac{\phi \mu^{p-1}}{1 + \alpha} \right) \implies \lambda = \frac{\mu^{2-p}}{\phi} \left( \frac{1 + \alpha}{\alpha} \right) μ = λ α ( 1 + α ϕ μ p − 1 ) ⟹ λ = ϕ μ 2 − p ( α 1 + α )
By matching the power dependencies of the cumulant function within the Exponential Dispersion Family, the relationship between p p p and the Gamma shape α \alpha α is established as: p = α + 2 α + 1 p = \frac{\alpha + 2}{\alpha + 1} p = α + 1 α + 2
Solving for the Gamma shape parameter α \alpha α : α ( p − 1 ) = 2 − p ⟹ α = 2 − p p − 1 \alpha(p - 1) = 2 - p \implies \alpha = \frac{2 - p}{p - 1} α ( p − 1 ) = 2 − p ⟹ α = p − 1 2 − p
Because 1 < p < 2 1 < p < 2 1 < p < 2 , both ( 2 − p ) > 0 (2 - p) > 0 ( 2 − p ) > 0 and ( p − 1 ) > 0 (p - 1) > 0 ( p − 1 ) > 0 , ensuring α > 0 \alpha > 0 α > 0 .
Substituting α \alpha α into the expressions for λ \lambda λ and θ \theta θ yields the fundamental actuarial bridge formulas: Gamma Shape: α = 2 − p p − 1 \textbf{Gamma Shape:} \quad \alpha = \frac{2 - p}{p - 1} Gamma Shape: α = p − 1 2 − p P o i s s o n M e a n : λ = μ 2 − p ϕ ( 2 − p ) \mathbf{Poisson\;Mean:} \quad \lambda = \frac{\mu^{2-p}}{\phi (2 - p)} Poisson Mean : λ = ϕ ( 2 − p ) μ 2 − p Gamma Scale: θ = ϕ ( p − 1 ) μ p − 1 \textbf{Gamma Scale:} \quad \theta = \phi (p - 1) \mu^{p-1} Gamma Scale: θ = ϕ ( p − 1 ) μ p − 1
Actuarial Implication for Zero-Claim Probability
Substituting λ \lambda λ into the point mass at zero yields: P ( Y = 0 ) = exp ( − λ ) = exp ( − μ 2 − p ϕ ( 2 − p ) ) P(Y = 0) = \exp(-\lambda) = \exp\left( -\frac{\mu^{2-p}}{\phi (2 - p)} \right) P ( Y = 0 ) = exp ( − λ ) = exp ( − ϕ ( 2 − p ) μ 2 − p )
Because 2 − p > 0 2 - p > 0 2 − p > 0 , as expected pure premium μ \mu μ increases, the exponent becomes more negative, causing P ( Y = 0 ) P(Y = 0) P ( Y = 0 ) to decrease. Higher-risk policyholder segments have a substantially lower probability of remaining claim-free, precisely matching actuarial observations.
Direct Pure Premium vs. Two-Stage Frequency-Severity Modeling
When pricing personal or commercial lines, casualty actuaries must decide between fitting a single direct Tweedie GLM or building a two-stage frequency-severity architecture. PURE PREMIUM MODELING PARADIGMS
- │
- ┌───────────────────────┴───────────────────────┐
- ▼ ▼
- DIRECT TWEEDIE MODEL TWO-STAGE ARCHITECTURE
- Response: Loss / Exposure Stage 1: Frequency GLM (Poisson)
- Prior Weights: Exposure Stage 2: Severity GLM (Gamma/IG)
- Link: Log Pure Premium = Frequency * Severity
- Single Objective Optimization Multiplicative Combination
- Comprehensive Actuarial Comparison
Evaluation Dimension Direct Tweedie GLM ( 1 < p < 2 1 < p < 2 1 < p < 2 ) Two-Stage (Poisson Frequency × \times × Gamma Severity)
Model Objective Optimizes pure premium deviance directly in one step. Optimizes frequency and severity deviances independently.
Target Variable Loss cost Y i = Loss i / e i Y_i = \text{Loss}_i / e_i Y i = Loss i / e i across all policies. Stage 1: Counts N i N_i N i ; Stage 2: Claim severity Z i > 0 Z_i > 0 Z i > 0 .
Prior Weights & Offsets Prior weight ω i = e i \omega_i = e_i ω i = e i (earned exposure); no offset. Stage 1: Offset ln ( e i ) \ln(e_i) ln ( e i ) ; Stage 2: Prior weight ω i = N i \omega_i = N_i ω i = N i .
Underwriting Insights Blurs frequency and severity drivers into a single net relativity. Explicitly distinguishes whether high cost is driven by frequency (e.g., mileage) or severity (e.g., luxury car parts).
Varying Policy Limits & Deductibles Difficult to adjust; severity limits distort aggregate pure premium. Straightforward; deductibles and policy limits apply cleanly to the conditional severity distribution.
Zero-Claim Handling Natively accommodates zero-claim policies via point mass e − λ e^{-\lambda} e − λ . Excludes zero-claim policies entirely from the severity stage.
Operational Overhead 1 model pipeline, 1 set of rate filing exhibits, 1 monitoring report. 2 model pipelines, 2 sets of diagnostics, combined tracking complexity.
Sample Size Disparity Trained on all policies ( n = 1 , 000 , 000 n = 1,000,000 n = 1 , 000 , 000 ). Frequency on all policies ( n = 1 , 000 , 000 n = 1,000,000 n = 1 , 000 , 000 ); Severity on claims only ( n = 80 , 000 n = 80,000 n = 80 , 000 ).
Strategic Guidance for Exam and Practice
Use Direct Tweedie when : Speed to market, operational simplicity, and direct loss cost optimization are paramount, particularly in high-volume personal lines (personal auto comprehensive, renters) with uniform policy limits.
Use Two-Stage Models when : Analyzing lines with high limit variability, significant deductible options (commercial property), or when underwriters require separate frequency and severity relativities to design underwriting guidelines.
Estimating and Profiling the Index Parameter p p p
In standard Generalized Linear Models (such as Poisson or Gamma), the variance power p p p is known ( p = 1 p = 1 p = 1 or p = 2 p = 2 p = 2 ), allowing the parameters β \boldsymbol{\beta} β to be solved directly via Iteratively Reweighted Least Squares (IRLS). In Tweedie models, however, the power parameter p p p is unknown and enters the variance function nonlinearly ( V ( μ ) = μ p V(\mu) = \mu^p V ( μ ) = μ p ).
Because p p p determines the Exponential Dispersion Family density itself, p p p cannot be estimated simultaneously with β \boldsymbol{\beta} β using standard IRLS. Instead, actuaries employ Profile Log-Likelihood Optimization .
- The Profile Log-Likelihood Algorithm PROFILE LIKELIHOOD ALGORITHM
- │
- ┌────────────────────────────┴────────────────────────────┐
- ▼ ▼
- 1. Define Grid for p 2. Loop over each p_k
- p in [1.05, 1.95], step = 0.05 Fit Tweedie GLM via IRLS
- Obtain beta_hat(p_k), phi_hat(p_k)
- │
- ┌─────────────────────────────────────────────────────────┘
- ▼
- 3. Calculate Exact Log-Likelihood
- Evaluate l_profile(p_k) using Dunn & Smyth series expansion
- │
- ▼
- 4. Identify Optimal p*
- p* = argmax l_profile(p_k)
- Refine via golden-section search near p*
Define Search Interval : Construct a discrete grid over the theoretical domain p ∈ ( 1.01 , 1.99 ) p \in (1.01, 1.99) p ∈ ( 1.01 , 1.99 ) (e.g., p ∈ { 1.1 , 1.2 , … , 1.9 } p \in \{1.1, 1.2, \dots, 1.9\} p ∈ { 1.1 , 1.2 , … , 1.9 } , refined to step size 0.01 0.01 0.01 ).
Conditional IRLS Fit : For each fixed candidate value p k p_k p k :
Fit the GLM for β \boldsymbol{\beta} β using standard IRLS with variance function V ( μ ) = μ p k V(\mu) = \mu^{p_k} V ( μ ) = μ p k .
Estimate the dispersion parameter ϕ ^ ( p k ) \hat{\phi}(p_k) ϕ ^ ( p k ) from the Pearson statistic.
Evaluate Exact Log-Likelihood : Evaluate the full log-likelihood ℓ ( β ^ ( p k ) , ϕ ^ ( p k ) ; p k ) \ell(\hat{\boldsymbol{\beta}}(p_k), \hat{\phi}(p_k); p_k) ℓ ( β ^ ( p k ) , ϕ ^ ( p k ) ; p k ) . Because the Tweedie normalizing function c ( y , ϕ ) c(y, \phi) c ( y , ϕ ) has no closed-form analytic expression for 1 < p < 2 1 < p < 2 1 < p < 2 , software evaluates the infinite series representation using the Dunn and Smyth (2005) Fourier inversion algorithm .
Maximize Profile Curve : Plot the profile log-likelihood curve against p p p and select the optimal value: p ∗ = arg max p ∈ ( 1 , 2 ) ℓ profile ( p ) p^* = \arg\max_{p \in (1, 2)} \ell_{\text{profile}}(p) p ∗ = ar g p ∈ ( 1 , 2 ) max ℓ profile ( p )
Actuarial Interpretation of the Optimal p p p
Lower values ( p → 1.1 − 1.3 p \to 1.1 - 1.3 p → 1.1 − 1.3 ) : Indicate that pure premium volatility is dominated by claim frequency . Claims occur with Poisson-like variation, while individual claim severities are relatively constant (low variance). Typical for auto glass or roadside assistance.
Higher values ( p → 1.7 − 1.9 p \to 1.7 - 1.9 p → 1.7 − 1.9 ) : Indicate that pure premium volatility is dominated by claim severity . Claim frequency is low, but loss sizes exhibit extreme right-skewness and high variance. Typical for commercial liability and excess property.
- Software Implementation Reference
- R ( tweedie and cplm packages) : library(tweedie)
- # Step 1: Profile p over a grid
- p_prof <- tweedie.profile(pure_premium ~ age + territory,
- weights = exposure, data = df,
- p.vec = seq(1.1, 1.9, by = 0.05))
- best_p <- p_prof$p.max
- # Step 2: Fit model with optimal p
- fit <- glm(pure_premium ~ age + territory,
- family = tweedie(var.power = best_p, link.power = 0),
- weights = exposure, data = df)
- Python ( statsmodels and scikit-learn ) : # scikit-learn TweedieRegressor with log link
- from sklearn.linear_model import TweedieRegressor
- reg = TweedieRegressor(power=1.5, link='log')
- reg.fit(X, y_pure_premium, sample_weight=df['exposure'])
- Actuarial Exam Traps & Pitfalls
Warning
Trap 1: Passing Exposure as an Offset Instead of Prior Weights in Tweedie In Poisson models ( p = 1 p = 1 p = 1 ), exposure enters as an offset ln ( e i ) \ln(e_i) ln ( e i ) because V ( μ ) = μ V(\mu) = \mu V ( μ ) = μ . In Tweedie models ( 1 < p < 2 1 < p < 2 1 < p < 2 ), loss cost Y i = Loss i / e i Y_i = \text{Loss}_i / e_i Y i = Loss i / e i is modeled directly, and exposure must enter as prior weights ω i = e i \omega_i = e_i ω i = e i , NOT as an offset. Setting offset = log(exposure) in a Tweedie model distorts the variance function and corrupts parameter estimates.
Warning
Trap 2: Specifying p ≤ 1 p \le 1 p ≤ 1 or p ≥ 2 p \ge 2 p ≥ 2 for Pure Premium The Compound Poisson-Gamma distribution with a point mass at zero exists strictly on the open interval 1 < p < 2 1 < p < 2 1 < p < 2 . Setting p = 1 p = 1 p = 1 reverts to a Poisson count model (incapable of modeling continuous loss amounts), while p = 2 p = 2 p = 2 reverts to a Gamma severity model (incapable of modeling zero losses). Any code setting p p p outside ( 1 , 2 ) (1, 2) ( 1 , 2 ) will fail to model pure premium correctly.
Warning
Trap 3: Believing Tweedie Provides Separate Frequency and Severity Relativities A direct Tweedie GLM outputs a single set of rating coefficients β \boldsymbol{\beta} β that represent the net effect on pure premium. It cannot tell an underwriter whether an adverse relativity (e.g., factor 1.35 1.35 1.35 ) is caused by high frequency or severe shock losses. If separate relativities are legally or operationally required, a two-stage model must be deployed.
Warning
Trap 4: Attempting to Optimize p p p Inside Standard IRLS Candidates often assume statistical software estimates p p p automatically via iteratively reweighted least squares. IRLS requires the variance function to be fixed. The index parameter p p p must be tuned via profile log-likelihood or out-of-fold cross-validation.