Overview of the specialty
Actuarial predictive analytics estimates the expected value of a risk variable for each policy or customer from its characteristics: number of claims, cost of each claim, probability of cancellation, probability that a claim is fraudulent, or the ultimate amount of an open claim. The difference from descriptive statistics is the purpose: the model is used to decide, policy by policy, what price to charge, which risk to accept, which claim to investigate or which customer to retain.
The problem it solves is adverse selection. If an insurer charges the same premium to drivers with very different risk, the good risks leave for the competitor that identifies them better and the portfolio is left with the bad ones. That is why segmentation is not a commercial luxury but a condition of rate adequacy: the LISF (Insurance and Surety Institutions Law) requires net risk premiums to be determined on technical bases, and the best technical basis available today is a well-validated multivariate model.
The generalized linear model (GLM) has been the industry standard since the 1990s because it handles skewed distributions (Poisson for counts, Gamma for amounts, Tweedie for pure premium), isolates the effect of each variable while controlling for the others and produces multiplicative factors that translate directly into a rate. Gradient boosting models (GBMs) gain accuracy by capturing nonlinearities and interactions, and today they coexist with the GLM: the GBM discovers the structure and the GLM formalizes it.
In Mexico the specialty is most visible in auto, major medical and homeowners insurance, where there is enough data volume, and increasingly in life insurance for simplified underwriting. External sources expand the set of variables: vehicle usage telematics, geospatial data on flood, earthquake and crime by postal code, and industry data. The actuary is the one who ensures that predictive power becomes a rate that is adequate, fair and documented in the technical note.
Key functions
Frequency and severity pricing
Models separately the number of claims per unit of exposure (Poisson or negative binomial) and the average cost (Gamma or lognormal), or the pure premium directly with Tweedie. Converts the coefficients into relativities by territory, age, use, vehicle or sum insured, and builds them into the rate together with expense and profit loadings.
Predictive underwriting
Builds scores that sort applications into automatic acceptance, manual review or decline, especially in major medical and simplified-issue life insurance. Measures the effect on loss experience of replacing medical exams with data and sets thresholds that balance issue time and risk selection.
Cancellation (churn) prediction
Estimates the probability of non-renewal with logistic regression, GBMs or survival models, using tenure, loss experience, premium increase, channel and payment behavior. The result feeds retention campaigns and the lapse assumption in long-term products.
Fraud detection
Combines supervised models trained on confirmed cases, unsupervised anomaly techniques and network analysis linking policyholders, repair shops, hospitals and adjusters. Prioritizes which claims to refer for investigation based on fraud probability times amount at risk.
Enrichment with telematics and geospatial data
Incorporates driving variables (mileage, time of day, harsh braking, speed) and location variables (flood or earthquake exposure, road density, theft by postal code). Smooths territories with thin experience spatially so the rate is not built on noise.
Reserve prediction and ultimate claim cost
Estimates the ultimate cost of each open claim with individual models (characteristics of the claim, the injured party and the provider) that complement aggregate triangle methods. Useful for detecting early deviations and for testing the adequacy of technical reserves.
Techniques, models and formulas
The typical workflow is: define the response variable and the exposure, build the dataset, fit a benchmark GLM, test a GBM to identify interactions and nonlinearities, carry the improvements into the final model, validate out of sample with segmentation and calibration metrics, and translate the result into rating factors. Metrics are chosen according to use: deviance and calibration for the overall level, Gini and lift for ranking.
Poisson, Gamma and Tweedie GLMs
With a log link, effects are multiplicative and exposure enters as an offset. Poisson models frequency, Gamma models severity (variance proportional to the square of the mean) and Tweedie with power p between 1 and 2 models pure premium with a point mass at zero, useful when you do not want to separate frequency and severity.
Gradient boosting (GBM)
Sequential ensembles of shallow trees with Poisson, Gamma or Tweedie loss. They outperform the GLM when there are complex interactions; their drawback is interpretability, which is addressed with SHAP and monotonicity constraints (for example, that the premium does not fall as the sum insured rises).
Segmentation metrics: Gini, lift and AUC
The ordered Gini measures how much the model concentrates losses in the risks it predicts to be worst. Lift compares the observed loss experience of the riskiest decile with the least risky one; the double lift compares two models across the deciles of their ratio of predictions. AUC summarizes discriminating power in binary problems such as churn or fraud.
Survival models for churn
The Cox proportional hazards model or discrete-time models estimate the probability of cancellation in each period by tenure and covariates. They are better than a binary classifier when the time to cancellation matters, as in the lapse assumption for life insurance.
Spatial smoothing and geospatial data
Postal codes with little exposure are smoothed using information from neighbors (CAR models, kriging or spatial credibility). External hazard layers for flood, earthquake or crime are joined by coordinates and add signal where the company's own experience is scarce.
Reserving with individual models
Instead of aggregating claims in a triangle, the development of each claim is modeled from its characteristics. These models complement chain ladder and Bornhuetter-Ferguson, especially when the claims mix or the speed of payment changes.
Ni is the number of claims for policy i, ei its exposure in years (offset), xij its variables and βj the coefficients. Each eβj is a rating relativity.
μ is the mean and φ the dispersion parameter. With p = 1 you get Poisson and with p = 2 Gamma; between those values the distribution is compound Poisson with Gamma severity, with a point mass at zero and a continuous tail, exactly like the pure premium of a policy.
Policies are sorted from lowest to highest predicted risk and grouped into K groups; Xk is the cumulative share of exposure and Yk the cumulative share of losses. A model with no segmentation power gives G = 0.
Regulatory framework in Mexico
LISF art. 200 · Technical bases of the premium
Insurers must determine net risk premiums on technical bases so as to guarantee, with a high degree of certainty, that they can meet their obligations, and bonuses or dividends cannot reduce the net risk premium. A predictive model is the technical basis; commercial discounts are funded from loadings, not from the risk premium.
LISF art. 201 · Technical note signed by an actuary
The technical note must describe the coverage and the risks, the actuarial procedures for determining premiums and extra premiums and the technical justification of their adequacy, and it must be prepared by an actuary with a cédula profesional (professional license) and current certification. If the rate comes from a GLM or GBM, the procedure, the variables and the validation must be describable there.
CUSF Chapter 4.1 · Product filing
Regulates the filing of insurance products with the CNSF, including the contractual documentation and the technical note. Significant rate changes resulting from a new model mean checking whether the filed product still reflects the actual calculation procedure.
CONAC · Gross premium standard of practice
The CONAC (Colegio Nacional de Actuarios) Actuarial Standard of Practice on gross premiums for short-term insurance sets out how to support rate adequacy. A model with a high Gini but poorly calibrated in level does not comply: the standard looks at adequacy as well as segmentation. See Standards.
LFPDPPP 2025 · Telematics and geolocation
Location and driving data are personal data: they require a privacy notice with clear purposes and consent. The law recognizes the right to object to automated processing without human intervention that significantly affects the data subject, which means human review must be built into automated underwriting.
NAIC · Review of predictive models in rate filings
The white paper of the NAIC's actuarial task force (2020) describes what information U.S. regulators review for a GLM used in a rate filing: data, variables, validation and impact on policyholders. It is a useful guide for documenting models even when the supervisor is the CNSF.
Case study: premium for an auto profile with a GLM and Gini validation
An insurer fits a Poisson frequency GLM and a Gamma severity GLM for auto physical damage coverage. The base level is 0.065 claims per vehicle year with an average cost of $38,000. The goal is to price an 18 to 25 year old driver in a metropolitan area with a telematics device that rates their driving as good. Frequency relativities: territory 1.35, age 1.60, telematics 0.85. Severity relativities: territory 1.10, age 1.05. Loadings: acquisition 20%, administration 8% and profit 5% of the gross premium. Figures are illustrative.
- Expected frequency: 0.065 × 1.35 × 1.60 × 0.85 = 0.11934 claims per year. On the log scale this is equivalent to adding coefficients: ln 0.065 = -2.733, plus 0.300 (territory), 0.470 (age) and -0.163 (telematics), equal to -2.126.
- Expected severity: $38,000 × 1.10 × 1.05 = $43,890 per claim.
- Pure premium (risk premium): 0.11934 × $43,890 = $5,237.83 per year.
- Gross premium: loadings total 33%, so divide by 1 - 0.33 = 0.67. Result: $5,237.83 / 0.67 = $7,817.66 before policy fees and VAT.
- Model validation on an out-of-time sample: when policies are sorted by predicted risk into quintiles of equal exposure, observed losses are distributed 8%, 14%, 19%, 24% and 35%. The cumulative shares are 0.08, 0.22, 0.41, 0.65 and 1.00; the area under the Lorenz curve is 0.372, and the Gini is 1 - 2 × 0.372 = 0.256. The riskiest quintile concentrates 4.4 times the losses of the least risky one.
- Calibration review: in addition to the Gini, predicted frequency is compared with observed frequency in each quintile and by territory. If the total predicted differs from experience, the intercept is adjusted before the relativities are carried into the technical note.
Where they work and how they grow
Prepares exposure and claims datasets, fits GLMs under supervision, calculates relativities and produces loss experience reports by segment. Learns to spot data problems that distort a model.
Owns a frequency, severity, churn or fraud model end to end, including the challenger GBM, validation and the rate proposal. Presents the expected effect on loss experience and retention to underwriting and management.
Coordinates the model agenda for a line of business, decides which variables and external sources are incorporated, signs or reviews technical notes and monitors the gap between expected and actual loss experience.
Sets the institution's pricing and risk selection strategy, balances growth and profitability, and answers to the board and the CNSF for rate adequacy.
- Auto, major medical and property insurers (pricing and underwriting departments)
- Life insurers with simplified or digital underwriting
- Reinsurers that support ceding companies with pricing models
- Actuarial and analytics consultancies (Big Four and specialized firms)
- Insurtechs offering pay-per-kilometer or usage-based auto insurance
- Providers of telematics, geospatial and risk data
- Banks and lenders (credit risk and collections models)
Skills and tools
Technical
Tools
Personal
Certification and training
CAS Institute · CSPA
The Certified Specialist in Predictive Analytics combines property and casualty fundamentals, data and visualization, predictive modeling, an applied project and ethics. It is the credential closest to the day-to-day work of model-based pricing.
SOA · SRM, PA and ATPA exams
The Society of Actuaries' path to ASA includes statistics for risk modeling (SRM), predictive analytics (PA) and advanced topics in predictive analytics (ATPA), together with the Data Science for Actuaries micro-credential.
CONAC certification in Property and Casualty or Life and Health
Required to sign the technical note that files the rate with the CNSF. See Certification.
Graduate degree in statistics or data science
A master's degree in statistics, data science or actuarial science goes deeper into generalized linear models, statistical learning and Bayesian methods.
Trends reshaping the field
From the pure GLM to the augmented GLM
Teams use the GBM to detect interactions and nonlinearities and carry them into a GLM or a GAM that can be filed. Specialized platforms automate that translation with monotonicity constraints and smoothing.
Usage-based insurance
Telematics makes it possible to price by kilometers driven and driving behavior. The challenges are low initial uptake, the credibility of just a few months of data and consent for location data.
Climate risk in the rate
Geospatial layers for flood, hail and extreme heat are being built into homeowners and auto pricing, connecting predictive analytics with catastrophe models.
Organized fraud and network analysis
Insurance fraud is organized in networks of recurring participants. Graph models connecting policyholders, providers and adjusters detect patterns that a claim-level model cannot see.
Frequently asked questions
Why is the GLM still relevant if gradient boosting predicts better?
What is the difference between Gini and AUC?
Can gender or age be used in rating in Mexico?
How do I get started in predictive analytics as a student?
Go deeper
- Ley de Instituciones de Seguros y de Fianzas (Insurance and Surety Institutions Law), arts. 200 to 202 (Chamber of Deputies)
- CNSF, CUSF Chapter 4.1: Filing of insurance products
- Goldburd, Khare, Tevet and Guller, Generalized Linear Models for Insurance Rating, CAS Monograph No. 5
- Frees, Meyers and Cummings, Summarizing Insurance Scores Using a Gini Index, JASA (2011)
- NAIC CASTF, Regulatory Review of Predictive Models (2020)
- Ley Federal de Protección de Datos Personales en Posesión de los Particulares (Federal Law on the Protection of Personal Data Held by Private Parties)
- The CAS Institute, Certified Specialist in Predictive Analytics