Gains:
- Ability to establish generalized linear model (GLM), risk factor, tariff and risk premium concepts with the support of artificial intelligence and translate the coefficients into business language
- Ability to discuss the balance of variable selection, interaction, overfitting, and interpretability of machine learning models (GBM) with artificial intelligence
- Being able to understand that the pricing model is free from prohibited/discriminatory variables and the compliance of the final tariff with legislation and competition is left to the actuary.
The price of an insurance policy is not determined randomly. The risk of a traffic accident for a 22-year-old, newly licensed driver living in a big city is statistically higher than that of a 45-year-old driver with 20 years of experience; In a fair system, premiums should also be different. The actuary's job of measuring this difference and reflecting it on the price is called pricing and risk classification. In this unit, we will discuss the generalized linear model (GLM), the actuarial backbone of pricing, and its machine learning alternatives, and see how to safely use AI in this process.
A few basic terms. Risk factor is the variable that affects the risk and is used in pricing (age, vehicle age, region, intended use). The tariff is a set of rules that determine how the premium will be calculated according to risk factors. The risk premium is the pure expected loss cost; When expenses, commissions, profit margins and capital costs are added, the commercial premium (sales price) is formed. Let's remind from the beginning: AI suggests variables, builds models, explains coefficients; But which variable is legal and ethical and whether the final tariff complies with legislation and competition belongs to the actuary and the compliance unit.
Why GLM is the standard in actuarial
The most used method in pricing is GLM. GLM stands for “generalized linear model”: it extends classical linear regression to fit non-normally distributed targets, such as damage data. It has two basic components. Distribution family: Poisson is chosen for frequency, gamma is chosen for intensity (logic in unit 2). Link function (link): usually logarithmic, which allows the factors to have a multiplicative effect. The multiplicative structure is very valuable in actuarial because that is exactly how the tariff is set up: base premium × age factor × region factor × vehicle factor.
The biggest advantage of GLM is interpretability. A coefficient tells you directly: "the young driver group increases the frequency by 1.6 times compared to the reference group." This transparency is critical in three ways: (1) the regulator and internal audit can understand and approve the model; (2) a prohibited/discriminatory effect is visible; (3) price logic can be explained to the sales and management team. More complex machine learning models (below) sometimes give higher accuracy, but at the expense of transparency.
Hint: GLM coefficient is on log scale. Ask the AI to convert the coefficient into a "multiplying factor" with exp(coefficient); It would be much easier to translate into business language (e.g. coefficient 0.47 → factor ≈ 1.60 → “60% riskier”).
Variable selection, overfitting and machine learning
The most critical decision in pricing is which variables will remain in the model. There are two traps. Few variables blind the model: if the important risk driver is missed, the price will be wrong. Too variable / overfitting makes the model memorize past data: the model mistakes noise for signal and fails with new policies. Balance is established through cross-validation — dividing the data into training and test pieces and testing it on data that the model does not see, simplicity and actuarial reasoning.
Interaction is also important: sometimes the combined effect of two factors differs from the sum of their separate effects (a young + sporty vehicle may be riskier than the sum of their individual effects). In GLM, interactions are added explicitly.
In recent years, models such as GBM (gradient boosting machine — a powerful machine learning method that combines many small decision trees) have become common in pricing. These automatically capture complex patterns and interactions, often giving more accurate predictions than GLM. But it comes at a price: the tendency to be a "black box". Common practice today is to use GBM for discovery and idea generation and GLM for the final transparent tariff; or interpreting the GBM output with explainability tools such as SHAP.
The following table compares the two approaches:
criterion
GLM
GBM (machine learning)
Interpretability
High (coefficient on)
Low (requires annotation tool)
Interaction capture
Added manually
automatic
Risk of overfitting
low-medium
High (careful adjustment required)
Regulator/audit approval
easy
Difficult, requires additional documentation
Typical usage
final tariff
discovery, comparison
How to use AI in pricing
1) Translating the GLM coefficient into business language:
I set up a frequency GLM (Poisson, log link). Some coefficients (log scale): age_group_18_25: 0.47 ; region_majority: 0.22 ; arac_yasi_10plus: -0.15. Convert each one into a multiplier factor with exp() and explain it in one sentence that a manager can understand. Also remind what the reference group is.
2) Variable/interaction discussion:
Your role: pricing assistant. My candidate variables for the traffic frequency model are: age, gender, region, vehicle age, engine power, annual km, profession. - Which variables might be risky from a regulatory/ethical perspective (e.g. indirect discrimination)? - Which bilateral interactions would make sense to try? - What verification steps should I follow to avoid overfitting? Decision making; Give me the framework for control and discussion.
3) Overfitting control code:
Write commented code that evaluates a GLM with k-fold cross-validation using Python + scikit-learn / statsmodels. Use Poisson deviation as the metric. Compare the training and test scores and explain in the comments line how to read the overfitting sign. The library is fake.
4) Discrimination/surrogate variable screening:
'Postcode' turned out to be a strong variable in my pricing model. Explain the risk of this indirectly representing a forbidden characteristic (e.g. ethnicity, income) as a proxy variable. - What analysis should I do to test for this risk? - What alternatives are there to reduce the risk? Providing legal advice; draw a technical and ethical framework.
Weak prompt / Strong prompt
Weak prompt:
Establish the best pricing model for traffic insurance.
"Best" is undefined; No data, no restrictions, no legislation. AI gives a generic, inapplicable answer.
Powerful prompt:
Your role: assistant to the pricing actuary. Context: I am building a traffic branch frequency model, GLM (Poisson, log link). The variables I have: age group, vehicle age, region (province), annual km, intended use. Constraint: gender cannot be used by legislation; I need to avoid indirect discrimination.Task:1) Suggest initial model specification with these variables (including reference groups).2) Give reasons for 2 interactions I should try.3) Give list of steps to check overfitting.4) Add special note of attention for 'region' in terms of indirect discrimination.I will make the decision; You provide the framework and justification.
three mini cases
Case 1 — The value of transparency. A company presented a very accurate pricing model it built with GBM to the regulator, but could not explain the coefficients; approval delayed. The actuary constructed a GLM that captured the same signals; Accuracy dropped by 2 percent, but because every factor could be accounted for, the model was validated within a month. GBM was saved for reconnaissance, GLM became tariff. YZ quickly produced the code and regulator description comparing the performance of the two models.
Case 2 — The overfitting trap. A helper got a perfect score on training data when he added more than 40 variables and numerous interactions to the model. But on cross-validation, the test score collapsed: the model had memorized the noise. Real-world performance increased when the number of variables was reduced to 12 and simplified. Lesson: look at the unprecedented data score, not the training score.
Case 3 — Indirect discrimination. In one model, the "neighbourhood" variable explained the premium strongly. The analysis showed that this variable largely overlapped with ethnic concentration, i.e., it was a proxy for a forbidden characteristic. The actuary replaced this variable with more neutral risk indicators (road type, parking condition); both ethical and legal risks have decreased. AI produced correlation analysis measuring overlap and alternative variable suggestions; The actuary and compliance unit made the decision.
Common mistakes
- Making the uninterpretable model the ultimate recipe. A price that cannot be explained to the regulator, audit and customer is unsustainable; Transparency is essential.
- Relying on education score. Overfitting is only detected in unseen data (cross-validation).
- Using forbidden/surrogate variables without checking. Variables such as zip code and occupation may carry indirect discrimination; Be sure to test it.
- Confusing risk premium with commercial premium. Pure damage cost alone is not the selling price; expense, profit and cost of capital are added.
- Leaving the coefficient on a log scale and misinterpreting it. Commenting without converting it to a multiplier factor with exp() will cause an error.
Caution: AI's recommendation of the model that "gives the best accuracy" is not automatically the model that "should be used". Transparency, fairness and compliance with legislation are mandatory criteria as well as accuracy in actuarial pricing.
In summary
Pricing is the process of measuring risk with risk factors and converting it into a fair premium. GLM is the standard of actuarial pricing with its multiplicative and interpretable structure; Coefficients are converted to factors with exp(). Machine learning models such as GBM may be more accurate, but they reduce transparency and are typically used for discovery. Variable selection should be protected against overfitting by cross-validation, and indirect discrimination (surrogate variable) should be carefully controlled. AI builds the model, writes code and explains the coefficient; But legal-ethical compliance and final tariff belong to the actuary and compliance unit.
Application task
List 5-6 risk factors for a major. Ask the AI for (a) an initial specification of the GLM with these factors, (b) 2 interactions to try and their rationale, (c) a screening of variables that might be at risk for indirect discrimination. Then, give an example of 3 log-coefficients, ask the AI to convert it into a multiplier factor with exp() and transfer it to the business language, and verify the factors manually.
checklist
- [ ] Can my model be interpreted; Can I translate the impact of each factor into business language?
- [ ] Have I checked for overfitting by cross-validation?
- [ ] Have I scanned for indirect discrimination for prohibited and proxy variables?
- [ ] Have I consciously separated the risk premium from the commercial premium?
- [ ] Did I convert the coefficients into multipliers with exp() and interpret them correctly?
- [ ] Did I make the final tariff decision together with the legislation and compliance unit?