Section 6 : Segmentation Flashcards
(175 cards)
1
Q
What is adverse selection and why is it problematic?
A
When higher-risk individuals are more likely to buy insurance due to asymmetric info, leading to losses if not priced appropriately.
βWhen improper classification causes loss of favorable risks and gain of unfavorable ones.β
π Source: Module 4, Section 4.6.1 (p.21β22)
2
Q
How can big data help mitigate adverse selection?
A
By enabling more granular risk segmentation and targeted pricing based on behavioral and external data.
π Source: Module 4, p.22 + Big Data Paper p.9
3
Q
Whatβs the risk of using overly granular pricing?
A
It may result in discrimination or regulatory concerns, especially if protected classes are disproportionately affected + probably less credible and sensible to big swings in the following years
π Source: Big Data Paper p.6, 13β14
4
Q
Whatβs the difference between risk and uncertainty in modeling?
A
Risk is quantifiable; uncertainty is not. Actuarial models work better when risks are measurable.
π Source: Big Data Paper p.7
5
Q
Why Zip Codes variable is useful but hard to model ethically.
A
useful in predicting loss, but can proxy for race or income, raising fairness issues.
π Source: Module 4, p.11β12 + Big Data Paper p.6
6
Q
What is the purpose of using external data in pricing models?
A
To enrich internal data, improve predictive power, and overcome limitations of sparse in-house data.
π Source: Module 4, Section 4.3.1 (p.11)
7
Q
What pricing response might you suggest if a competitor charges 10% less?
A
Consider segment-specific discounting, value-added services, or reviewing expense/reinsurance efficiency.
π Source: Module 4, Exercise 4.2 (p.10)
8
Q
How do actuaries handle missing or default data?
A
Imputation, modeling missingness (by replacing missing values by mean), or supplementing with external benchmarks.
π Source: Module 4, p.12
9
Q
How do you model rare but severe losses?
A
Use credibility weighting with industry data and fat-tailed distributions like Pareto.
π Source: Module 4, Section 4.3.4 (p.14)
10
Q
Why are GLMs popular in actuarial pricing?
A
They handle skewed distributions, provide interpretable outputs, and suit exposure-based data.
π Source: Module 5, Section 5.3 (p.9β15)
11
Q
When might GLMs be insufficient?
A
When relationships are highly nonlinear or involve complex interactionsβdata mining may be better.
π Source: Module 5, Section 5.4 (p.19β22)
12
Q
What is the role of PvO (Predicted vs Observed) charts?
A
To validate that model predictions align with actual outcomes across segments.
π Source: Module 4, Section 4.7.2 (p.24)
13
Q
What do lift charts assess?
A
The effectiveness of a model at discriminating between high and low-risk segments.
π Source: Module 4, Section 4.7.3 (p.25)
14
Q
Why might you choose a simpler model over a better-fitting one?
A
To enhance interpretability, regulatory acceptance, and robustness against overfitting
π Source: Module 5, Section 5.3.2 (p.12β13)
15
Q
What are non-risk factors in pricing decisions?
A
Competition, brand strategy, customer loyalty, marketing channels, and regulation.
π Source: Module 4, Section 4.8 (p.28)
16
Q
How can you use quote conversion data in pricing?
A
To model demand elasticity and optimize price points for profitability.
π Source: Module 4, Section 4.8.2 (p.30)
17
Q
When should you conduct pricing impact analysis?
A
Before launching a new pricing strategy to assess profit, growth, and risk implications.
π Source: Module 4, Section 4.8.5 (p.34)
18
Q
How should pricing approaches differ between stakeholders (e.g., consumers vs. regulators)?
A
Consumers prioritize affordability and fairness; regulators prioritize accessibility, transparency, and non-discrimination; companies prioritize sustainability and profitability.
π Source: Big Data Paper p.6β7
19
Q
When is pricing with technical premiums not enough?
A
When market dynamics, strategic goals, or regulatory constraints require deviation.
π Source: Module 4, p.7β8
20
Q
Why is exposure normalization key in experience studies?
A
It ensures fair comparison of loss rates across policies or time.
π Source: Module 4, Section 4.3.2 (p.12)
21
Q
Whatβs the difference between disparate impact and disparate treatment?
A
Impact = unintentional unfair outcomes; Treatment = intentional bias.
π Source: Big Data Paper p.12
22
Q
How should you approach using socio-economic data?
A
Carefullyβbalance predictive power with risk of bias and regulatory limits.
π Source: Big Data Paper p.40β43
23
Q
Why is explainability important in model deployment?
A
To meet regulatory expectations and build consumer trust.
π Source: Big Data Paper p.8β9
24
Q
What is the goal of simple tabular analysis in risk classification?
A
to explore patterns and relativities using summary statistics across one or two rating variables.
25
What's a key limitation of tabular analysis?
It does not consider correlations between explanatory variables.
26
What is the goal of GLM parameter estimation?
To test if predictors significantly impact the outcome (i.e., significantly β 0).
27
Which distributions are typically used in insurance GLMs?
Poisson for frequency, Gamma for severity.
28
What's the difference between an offset and weights in a GLM?
Offset adjusts the mean; weights adjust the importance of observations.
29
How is the predicted value retrieved in a GLM?
By applying the inverse of the link function.
30
Which metrics are used to compare GLMs?
AIC, BIC, Scaled Deviance, F-test, Chi-square
31
What does a small AIC or BIC indicate?
A better-fitting model with fewer parameters.
32
What is cross-validation?
A technique that trains and validates the model across multiple data splits.
33
Why is residual analysis important in GLMs?
To detect patterns that show model misfit.
34
What is distributional bias?
When exposures are unevenly distributed across classes, impacting relativity calculations.
35
Why might a one-way approach fail?
It doesnβt account for interactions between variables.
36
What is the goal of minimum bias methods?
To adjust class relativities considering all combinations of risk factors.
37
What error metric is used to compare one-way and minimum bias methods?
Absolute errorβlower is better.
38
What is the main advantage of Decision Trees?
Easy to interpret and visualize.
39
What's a key risk with Decision Trees and Random Forests?
Overfittingβmodels may capture noise instead of signal.
40
How do Gradient Boosting Machines (GBM) work?
By building trees sequentially, each correcting the previous.
41
Why use Random Forests?
For improved accuracy and flexibility across data types.
42
What are key ethical concerns in modeling?
Data quality, transparency, causality, and fairness.
43
What is a proxy variable?
A variable that mimics a disallowed or biased variable.
44
What's the difference between interpretability and explainability?
Interpretable models explain themselves; explainable models need support to be understood.
45
What is lift in model performance?
The model's ability to separate good and bad risks.
46
What does a ROC curve show?
The trade-off between true and false positives across thresholds.
47
What is AUROC?
Area Under ROC Curveβcloser to 1 = better model.
48
What does the Gini index measure in modeling?
The ability to rank risks, not profitability.
49
What are common transformations for non-linear variables?
Binning, polynomial terms, and piecewise linear functions.
50
Whatβs the purpose of interaction terms?
To model the combined effect of variables.
51
What causes multicollinearity in GLMs?
Strong correlation between predictors.
52
How do you detect aliasing?
Perfect correlation between predictorsβinstability in model.
53
What are the high-level model-building steps?
Set goals, collect data, model, validate, implement, maintain.
54
When is regularization used?
To handle many variables and prevent overfitting.
55
In classification techniques, What does precision measure?
The proportion of predicted positives that are actually positive.
56
Describe the following objectives with respect to risk classification:
(i) fair or equitable premium rates
Fair premium means the rates that reflect expected cost of the insured.
57
Describe the following objectives with respect to risk classification:
ii) socially adequate premium rates
Socially adequate premium means the rates that reflect the affordability of
the insurance considering each person's economic situation.
58
Describe how the following objectives could be contradictory when rating by age
for drivers aged 18, 40, and 80 in automobile insurance.
(i) fair or equitable premium rates
(ii) socially adequate premium rates
Age 18: Typically, a high premium for younger individuals, but this may not be viewed as socially adequate depending on the jurisdiction.
Age 40: Although premiums calibrated to be fair may be deemed socially adequate at this age, there may be subsidies given the socially adequate concerns for younger (18) and older (80) individuals, leaving these premiums no longer at fair prices.
Age 80: Fair premiums would tend to be higher than socially adequate premiums for this age group.
59
Describe two operational considerations that affect the practicality of designing
and maintaining a risk classification system.
Any two of the following are acceptable:
β’ Objectivity: Where possible, the evaluation of a risk characteristic should be factual and not judgmental.
β’ Cost: One should note that costs arise from obtaining, storing, and analyzing the data required for actuarial work supporting the risk classification system.
β’ Verifiablity: The risk characteristics used in a risk classification system should be reliable and conveniently verifiable. Characteristics such as age, gender, and occupation can, in general, be reliably measured.
60
What is the product hierarchy in insurance products?
A structure showing product categories from basic to complex, helping identify pricing needs at each level.
π Chapter 3.1 Monograph #5
61
ow do regulations affect pricing strategies?
Regulations can limit rating factors or require justification for differentials, influencing model selection and implementation.
π Chapter 8.3 Monograph no 5
62
Why is pricing for different levels of risk necessary?
It ensures that premiums are aligned with expected losses, reducing cross-subsidization.
π Chapter 2.1 Monograph 5
63
What happens if no risk-based pricing is used?
It may result in underpricing high-risk segments and overpricing low-risk ones, causing portfolio deterioration.
π Chapter 3.1 Monograph no 5
64
How can GLMs help with pricing risks more accurately?
They model multiple risk factors simultaneously, capturing interactions and nonlinearities.
π Chapter 2.1, 5.6 Monograph #5
65
What is simple tabular analysis in segmentation?
It uses cross-tabulations to explore the relationship between a single factor and outcomes.
π Chapter 3.4 Monograph #5
66
Name one limitation of simple tabular analysis.
It cannot control for confounding variables or detect interactions.
π Chapter 3.4 Monograph 5
67
What makes GLMs suitable for segmentation analysis?
They model relationships using multiple predictors with flexible error structures.
π Chapter 2.1, 5.1 Monograph 5
68
How do machine-learning models differ from GLMs?
They often have higher predictive power but may lack interpretability.
π Chapter 9.3 Monograph 5
69
What are neural networks in the context of pricing?
Machine-learning models that model complex, nonlinear relationships using interconnected nodes (neurons).
π Chapter 9.3 Monograph 5
70
What is data mining in insurance pricing?
The process of discovering patterns in large datasets to inform pricing strategies.
π Chapter 3.4, 4.4 Monograph 5
71
What characterizes non-parametric methods?
They make fewer assumptions about data distributions, allowing flexible modeling.
π Chapter 9.3 Monograph 5
72
What are the key assumptions of GLMs?
Linearity in the link-transformed means, independence, and correct distribution specification.
π Chapter 2.1.1, 2.1.2 Monograph 5
73
What is the role of credibility in pricing models?
It adjusts results from smaller segments toward overall averages to stabilize estimates.
π Chapter 6.4 Monograph 5
74
What does "off-balance" mean in pricing?
When the average premium changes due to the model, requiring adjustment to maintain overall premium income.
π Chapter 2.6 Monograph 5
75
What is a critical first step in segmentation analysis?
Validating input data for accuracy, completeness, and consistency.
π Chapter 4.1β4.2 Monograph 5
76
How can model outputs be validated?
By checking fit metrics, residual patterns, and stability on holdout datasets.
π Chapter 6.3, 7.1β7.3 Monograph 5
77
How is credibility weighting applied in segmentation?
It blends segment-specific and overall estimates to avoid overfitting to noise.
π Chapter 6.4 Monograph 5
78
What should be considered when selecting segmentation techniques?
Use case goals, interpretability, data volume, and regulatory context.
π Chapter 5.3, 8.3 Monograph 5
79
What link function is typically used for Poisson-distributed data in GLMs?
The logarithmic link function.
π Chapter 2.1.2, 2.7.2 Monograph 5
80
What is an example of multivariate modeling in pricing?
Including age, vehicle type, and location simultaneously in a model to predict claims frequency.
π Chapter 5.6 Monograph 5
81
What is the importance of selecting the correct error structure in GLMs?
It ensures the variance assumptions align with the data, improving fit and interpretability.
π Chapter 2.1.1, 2.7 Monograph 5
82
What do residual plots reveal in GLMs?
Patterns that suggest model misspecification or data issues.
π Chapter 6.3 Monograph 5
83
What are parameter estimates in GLMs?
Coefficients that quantify the effect of predictors on the response variable.
π Chapter 2.1.2, 2.3 Monograph 5
84
How do you identify model errors in GLM outputs?
By analyzing residuals, leverage, and deviance statistics.
π Chapter 6.3 Monograph 5
85
Why are natural hazards important in pricing?
They significantly influence the frequency and severity of claims in certain regions.
π Chapter 9.2 Monograph 5
86
How do insurers adapt models for regulated pricing environments?
By applying constraints, blending models with mandated rates, or using off-balance adjustments.
π Chapter 2.6, 8.3 Monograph 5
87
How is pricing linked to outstanding claims analysis?
Past claims development patterns inform expected future costs.
π Implied context, Chapter 1, 9.3 Monograph 5
88
What is exposure analysis?
Studying policyholder risk exposure over time to inform pricing and reserves.
π Chapter 2.6 Monograph 5
89
Why are economic assumptions important in pricing?
Inflation and interest rates affect the present value of long-tail claim liabilities.
π Chapter 9.3 Monograph 5
90
What is the purpose of using interaction terms in GLMs?
To model how the effect of one variable depends on another.
π Chapter 5.6 Monograph 5
91
What does overdispersion indicate in a Poisson GLM?
Variance exceeds the mean, suggesting a negative binomial model might be better.
π Chapter 2.7.2 Monograph 5
92
Whatβs the key difference between GLMs and decision trees?
GLMs are parametric and require assumptions; trees are non-parametric and more flexible.
π Chapter 9.3 Monograph 5
93
What is model parsimony?
The principle of using the simplest model that adequately fits the data.
π Chapter 6.2 Monograph 5
94
Why is interpretability important in pricing models?
It aids in regulatory approval, business understanding, and communication.
π Chapter 8.3 Monograph 5
95
What is the Akaike Information Criterion (AIC)?
A measure used to compare models, balancing fit and complexity.
π Chapter 6.2.2 Monograph 5
96
Why might a neural network be inappropriate for regulatory filings?
Due to its black-box nature, which limits transparency.
π Chapter 9.3 Monograph 5
97
Whatβs the role of exposure in frequency modeling?
It adjusts for time-at-risk differences among policies.
π Chapter 2.6 Monograph 5
98
How can we detect multicollinearity in a GLM?
Using variance inflation factors (VIFs).
π Chapter 2.9 Monograph 5
99
What is shrinkage in model estimation?
A technique to prevent overfitting by penalizing extreme coefficients.
π Chapter 9.3 Monograph 5
100
What are credibility adjustments in GLMs?
Blending base-level predictions with group-specific experience.
π Chapter 6.4 Monograph 5
101
What is a link function in GLMs?
A transformation that connects the linear predictor to the mean of the distribution.
π Chapter 2.1.2 Monograph 5
102
What metric helps evaluate prediction error for continuous outcomes?
Root Mean Squared Error (RMSE).
π Chapter 6.1, 7.1 Monograph 5
103
What is the difference between full credibility and partial credibility in rating?
Full credibility means enough data exists for a reliable estimate; partial credibility blends observed experience with other sources.
π Chapter 6.4 Monograph 5
104
When is it appropriate to apply external credibility (e.g., BΓΌhlmann-Straub) after a GLM is fit?
When model outputs need to be adjusted due to insufficient data at granular levels or to meet business/regulatory thresholds.
π Chapter 6.4 Monograph 5
105
What is the effect of including an interaction between a continuous and a categorical variable in a GLM?
It allows the slope of the continuous predictor to vary by category level.
π Chapter 5.6.2 Monograph 5
106
Why are interaction terms important for pricing?
They capture dependencies between predictors that affect expected losses in a nonlinear or subgroup-specific way.
π Chapter 5.6 Monograph 5
107
How can you test the significance of an interaction term in GLMs?
Use nested model comparison via deviance/F-test or check p-values for interaction terms.
π Chapter 6.2.1 Monograph 5
108
What is the difference between an offset and a weight in GLMs?
Offsets adjust the expected mean (systematic component), weights adjust the assumed variance (random component).
π Chapter 2.6 Monograph 5
109
In what scenario would a claim count model with an offset equal a frequency model with a weight?
When using a Poisson GLM: offset = log(exposure), weight = exposure (they are mathematically equivalent).
π Chapter 2.6 Monograph 5
110
Why is exposure not both an offset and a weight in a GLM?
Because doing so would improperly adjust both mean and variance, breaking the underlying assumptions.
π Chapter 2.6 Monograph 5
111
What does a lift chart show in model validation?
It compares the actual vs. predicted values across quantiles, showing the modelβs ability to differentiate risk.
π Chapter 7.2 Monograph 5
112
How does regularization (e.g., LASSO) prevent overfitting in GLMs?
By shrinking coefficients of less important variables toward zero, it reduces variance and model complexity.
π Chapter 10.5 Monograph 5
113
How can a GLM accommodate regulatory constraints like fixed base rates or territory factors?
By including these as offsets in the model to prevent re-estimation while accounting for their effect.
π Chapter 2.6, 8.3 Monograph 5
114
Why might a regulator reject a highly accurate model like a neural network?
Due to lack of transparency, interpretability, and challenges in justification of rates.
π Chapter 9.3, 8.3 Monograph 5
115
What does βrate relativityβ mean in regulatory filings?
The ratio of the premium for a given segment to the base rate; it must often be filed and justified.
π Chapter 2.4.2, 8.3 Monograph 5
116
How can model outputs be validated? (5 types of validations)
π 1. Residual Analysis
βͺ Deviance Residuals β Standardized measure of fit; large values may indicate outliers or poor fit (Ch. 6.3.1)
βͺ Working Residuals β Difference between observed and predicted, adjusted for variance; useful for diagnosing structure issues (Ch. 6.3.2)
βͺ Residual plots vs. fitted values or predictors help detect non-linearity, heteroscedasticity, or missing interactions.
π 2. Lift and Gini Metrics (Ch. 7.2)
βͺ Lift Charts β Sort risks by predicted value; plot actual vs. expected. Good models show steeper lift curves.
βͺ Gini Index β Measures rank-order discrimination power; a higher Gini means better ranking performance.
π 3. Double Lift & Loss Ratio Charts (Ch. 7.2.2, 7.2.3)
βͺ Double Lift β Compare actual-to-expected ratios across quantiles of both actual and predicted.
βͺ Loss Ratio Charts β Plot loss ratios (loss/premium) by predicted risk band. Flat lines near 1.0 are ideal.
π 4. ROC Curves for Binary Outcomes (Ch. 7.3.1)
βͺ Only for logistic models. Show tradeoff between true positive and false positive rate.
π 5. Cross-Validation (Ch. 4.3.4)
βͺ Split data into train/test or use k-fold. Assess fit and stability (consistency of results) across folds.
βͺ Avoid overfitting β a model that performs well on training but poorly on validation is overfit.
Source = Monograph 5
117
What is data leakage in cross-validation, and why is it dangerous?
Data leakage happens when information from outside the training data improperly influences the model, typically because it was unintentionally included during preprocessing or modeling.
π Examples from insurance modeling:
Including aggregated variables (e.g., policy-level claim ratios calculated using the full dataset) in training.
Splitting data after transformations like binning or scaling, causing knowledge of the test set to "leak" into training.
Preprocessing (e.g., imputation, scaling, binning) on the full dataset before splitting for cross-validation.
π Why itβs dangerous:
It creates artificially high model performance (low error or high lift), because the model βcheatsβ by learning patterns that wouldnβt exist in a real production environment.
When deployed, such models fail catastrophically on truly unseen data.
π This is discussed in Chapter 4.3.4, especially under cross-validation best practices. Monograph 5
118
What are the two main types of machine learning?
Supervised learning and unsupervised learning.
π Introduction, p.5 SOA
119
Why is insurance data particularly suited to machine learning methods?
It's often sparse, volatile, skewed, and time-variant.
π Introduction, p.5 SOA
120
What is the standard modeling method currently used in insurance pricing?
Generalized Linear Models (GLMs).
π Introduction, p.5 SOA
121
Name one key limitation of GLMs in insurance.
They assume a linear relationship (after transformation), which may not hold.
π Issues With GLM, p.7 SOA
122
How do GLMs handle credibility?
They offer no natural blending mechanism β only zero or full credibility.
π Issues With GLM, p.7 SOA
123
What is the goal of regularization?
To prevent overfitting by penalizing model complexity.
π Regularization, p.7 SOA
124
What does the LASSO penalty do?
Penalizes the absolute value of coefficients, encouraging sparsity.
π LASSO, p.7β8 SOA
125
What is the key difference between ridge and LASSO regression?
Ridge shrinks coefficients but doesnβt set any exactly to zero.
π LASSO vs Ridge, p.8 SOA
126
When is elastic net preferred over LASSO?
When predictors are highly correlated.
π Elastic Net, p.8 SOA
127
How is the regularization strength parameter (Ξ») usually chosen?
Through cross-validation.
π Regularization, p.8 SOA
128
What is subset selection in feature selection?
Evaluating all combinations of features and selecting the best based on AIC/BIC.
π Feature Selection, p.9 SOA
129
What is a major drawback of subset selection?
It becomes computationally infeasible for large p (number of predictors).
π Feature Selection, p.9 SOA
130
What is stepwise regression?
Iteratively adds/removes variables based on improvement in model fit.
π Feature Selection, p.9 SOA
131
What does Bayesian variable selection do differently?
It averages over all models weighted by their posterior probabilities.
π Bayesian Selection, p.9 SOA
132
What does CART stand for?
Classification and Regression Trees.
π CART, p.9 SOA
133
How does CART partition data?
Through sequential binary splits of predictor variables.
π CART, p.10 SOA
134
What are two limitations of CART?
High variance and lack of smoothness.
π CART Limitations, p.11 SOA
135
What is pruning in decision trees?
Removing less significant splits to prevent overfitting.
π CART, p.10 SOA
136
What is bagging?
Fitting multiple models to bootstrap samples and averaging predictions.
π Bagging, p.11 SOA
137
What does bagging reduce?
Variance.
π Bagging, p.11 SOA
138
What is the difference between bagging and random forests?
Random forests add a random subset of features at each split.
π Random Forests, p.12 SOA
139
140
Why do random forests outperform bagging?
Greater decorrelation between trees reduces ensemble variance.
π Random Forests, p.13 SOA
141
What is the main idea behind boosting?
Fit models sequentially to residuals from previous models.
π Boosting, p.14 SOA
142
What does the learning rate (Ξ») control in boosting?
How much each new model influences the final prediction.
π Gradient Boosting, p.15 SOA
143
What is the benefit of using shallow trees in boosting?
Prevents overfitting and keeps models interpretable.
π Boosting, p.15 SOA
144
What is a pseudo-residual in gradient boosting?
The gradient of the loss function with respect to the prediction.
π Gradient Boosting, p.15 SOA
145
What is the key innovation in BART?
Using a prior to limit the depth of each tree and encourage simplicity.
π BART, p.17 SOA
146
How does BART estimate models?
Using Bayesian backfitting via MCMC.
π BART, p.17
147
What does MARS stand for?
Multivariate Adaptive Regression Splines.
π MARS, p.18 SOA
148
What are basis functions in MARS?
Functions like max(x β c, 0) used to model nonlinear effects.
π MARS, p.18 SOA
149
How does MARS fit and refine its model?
Forward stepwise addition, then backward elimination.
π MARS, p.18 SOA
150
What is the biggest challenge of neural networks in insurance?
Lack of interpretability.
π Neural Nets, p.18 SOA
151
What is deep learning?
A multi-layer extension of neural nets that captures complex patterns.
π Neural Nets, p.18 SOA
152
What is clustering?
Grouping data points into segments with high internal similarity.
π Clustering, p.19 SOA
153
How can clustering help insurance modeling?
It reduces granularity, improves stability, and aids model convergence.
π Clustering, p.19 SOA
154
Name three clustering techniques.
K-means, DBSCAN, hierarchical clustering.
π Clustering, p.19 SOA
155
What metric is used for classification models?
AUC (Area Under the ROC Curve).
π Data section, p.6 SOA
156
What metric is used for regression models?
Mean Squared Error (MSE).
π Data section, p.6 SOA
157
What is cross-validation used for in ML?
Tuning hyperparameters and evaluating model performance.
π Regularization, p.8 SOA
158
What is the tradeoff between GLMs and ML methods?
GLMs are interpretable, ML methods are more flexible and accurate.
π Conclusion, p.21 SOA
159
What is a hyperparameter?
A tuning variable not estimated by the model (e.g., learning rate, tree depth).
π Boosting/Random Forests, p.15 SOA
160
What is variable importance in trees?
A metric indicating how much a variable contributes to model splits.
π Random Forests, p.14 SOA
161
Why might a highly accurate ML model not be used in practice?
It may be non-transparent or conflict with regulatory constraints.
π Interpretation, p.21 SOA
162
What is feature engineering in insurance?
Grouping raw variables into meaningful predictors like credit score.
π Feature Selection/Design, p.21 SOA
163
Which models outperformed GLMs in the SOA study?
Random forests, boosting, and BART β depending on the dataset.
π Conclusion, p.22 SOA
164
What makes ensemble models generally outperform single models?
They reduce variance and improve stability through averaging.
π Ensemble Methods, p.11β13 SOA
165
Why is model interpretability a key consideration in insurance?
To meet regulatory scrutiny and gain business acceptance.
π Interpretation, p.21 SOA
166
Why is stratified sampling useful in ML?
It ensures rare events (e.g., claims) are represented in training data.
π Random Forests, p.13 SOA
167
What is the typical split for training/testing data in the report
70% training, 30% testing.
π Regularization/Validation, p.8 SOA
168
Whatβs the key advantage of BART over traditional trees?
Built-in regularization via priors to control overfitting.
π BART, p.17 SOA
169
How do you choose the optimal number of trees in an ensemble?
Monitor test error; stop when performance plateaus.
π Bagging, p.11 SOA
170
Why is it hard to compare ML methods conclusively in insurance?
Lack of large, public benchmark datasets.
π Conclusion, p.22 SOA
171
What is the main risk of using very deep trees?
Overfitting to training data.
π CART, p.10 SOA
172
What does "shrinkage" mean in boosting?
Controlling the learning rate to avoid overfitting.
π Boosting, p.15 SOA
173
How does ensemble size (B) affect variance in bagging?
Larger B lowers variance but plateaus in effect.
π Bagging, p.11 SOA
174
What technique allows ML models to output prediction intervals?
Bayesian methods like BART.
π BART, p.17 SOA
175
Which method combines GLMs with tree-based approaches?
Random GLM or hybrid models.
π Additional Methods, p.20 SOA
