An auto insurance dataset has 12% of `driver_age` values missing, and the missingness is unrelated to age, claim amount, or any other variable. Which missing-data mechanism best describes this?
Data Prep Eda
Question · Easy
A workers' compensation dataset shows that older policyholders are systematically less likely to report `prior_claims_count` (it is missing more for ages 60+). Conditional on observed age, the missingness does not depend on the unobserved claim count itself. This pattern is best described as:
Question · Medium
Chapter 1: PCPA Exam & Project Orientation
1.1 PCPA Exam Format, Scoring & Registration
Study guide
A homeowners dataset has loss amounts that are heavily right-skewed with a long upper tail. Which transformation is most commonly applied before fitting a linear model on `loss_amount`?
Finance Exam Prep Podcast
Listen on Spotify
Podcast
Which transformation is the appropriate generalization of Box-Cox when the predictor or response can take zero or negative values?
An actuary uses the IQR method to flag outliers in `policy_premium`. The first and third quartiles are $400 and $1,200, respectively. The conventional outlier fences are at:
Which method detects outliers by isolating observations using random splits on randomly selected features, with shorter average path lengths indicating anomalies?
1.2 The PCPA Project: Windows, Deliverables & Grading
Capping the top 1% and bottom 1% of a continuous predictor at the 99th and 1st percentile values is called:
For an auto liability frequency model, the most appropriate exposure measure is generally:
Which exposure base is conventional for workers' compensation predictive modeling?
A product manager asks you to impute missing `vehicle_age` values using the mean of all observed ages. Which is the most important drawback to flag?
Chapter 2: Dealing with Data — Sourcing, Importing & Assessing
2.1 Requirements Gathering & Assessing Data Relevance
Multiple Imputation by Chained Equations (MICE) is most appropriate when:
Question · Hard
A categorical predictor `garage_type` has missing values for 8% of records. Of these, the most defensible imputation strategy without strong external information is typically:
In bivariate exploratory analysis for a P&C frequency model, which plot best reveals whether `driver_age` has a non-linear relationship with claim count per car-year?
When standardizing predictors before fitting a regularized GLM, the standardization should be:
2.2 Importing, Joining & Reshaping Data Sets
Which is the BEST measure of central tendency to report for a heavily right-skewed loss-amount distribution to communicate 'typical' loss to a non-technical audience?
For a pure premium (loss cost per exposure) model, the Tweedie distribution with power parameter p satisfying 1 < p < 2 is appropriate because:
Glm Pc Modeling
When fitting a Poisson GLM for claim frequency with claim count as the response, exposure should be incorporated as:
The canonical link function for the Gamma distribution in a GLM is:
2.3 Exploratory Data Analysis & Evaluating a Data Set
Which link function is most commonly used in practice for a Gamma severity GLM in P&C ratemaking, even though it is not the canonical link?
An actuary fits a frequency × severity model and a Tweedie pure-premium model on the same data. Which statement is generally correct?
Deviance in a GLM context is best described as:
Two nested Poisson GLMs are compared. Model A has 8 parameters with deviance 1,420; Model B (which adds 3 parameters) has deviance 1,402. The chi-squared 95th percentile for 3 degrees of freedom is 7.81. Which conclusion is correct at the 5% level?
2.4 Data Quality Under ASOP 23 & Modeling Under ASOP 56
AIC and BIC differ in their penalty for the number of parameters k. Which statement is true for a sample size n > 7?
A log-link Poisson GLM produces a coefficient of 0.20 for `urban_indicator` (1 = urban, 0 = rural). The predicted multiplicative effect on claim frequency for an urban policy is approximately:
When fitting a binomial GLM (e.g., probability of large loss > $50,000), the canonical link function is:
A Tweedie GLM with power p = 1.5 implies which of the following relationship between the variance and the mean?
Chapter 3: Dealing with Data — Conditioning P&C Data for a GLM
3.1 Exposure Bases, Weights & Data Grain
A homeowners severity GLM treats `roof_age` as a continuous predictor and `construction_type` as a 4-level factor. To allow severity to differ by construction type at every roof age, the modeler should:
Quasi-Poisson and negative binomial GLMs are both used to address which common issue with claim count data?
An actuary uses a Pearson residual plot vs. the linear predictor for a Gamma severity GLM. A clear funnel shape (residual variance increasing with the predictor) suggests:
Offsets and weights in a GLM differ in that:
3.2 Missing Data Mechanisms & Imputation Strategies
In a Gamma severity GLM with log link, an actuary fits separate models on `peril = wind` and `peril = water`. The most defensible reason to fit by peril rather than including peril as a factor with interactions is:
A Poisson GLM coefficient for `vehicle_symbol` (treated as continuous) is 0.05. Holding all else constant, a 4-symbol increase in `vehicle_symbol` changes predicted frequency by approximately:
When a GLM is fit on training data and then scored on a holdout, an unusually large drop in deviance lift between training and validation typically indicates:
In a Tweedie pure-premium GLM, which interpretation is correct for a coefficient of 0.30 on `urban_indicator` (log link)?
3.3 Outlier Diagnostics, Trimming & Winsorization
Which hyperparameter in XGBoost controls the L2 regularization on leaf weights?
Tree Based Pc
In a Random Forest, OOB (out-of-bag) error refers to:
For a regression tree (continuous response, e.g., loss amount), the standard splitting criterion is:
For a binary classification tree (e.g., large-loss flag), the most common splitting criterion is:
Loading more…