TL;DR
Predictive analytics in insurance uses policy, claims and outside data to score likely outcomes, so an insurer can price, route or act earlier. Five model families do most of the work: underwriting and pricing, claims triage and fraud scoring, lapse and retention, reserving, and catastrophe risk. Each one needs its own target, data, KPI and a named person who acts on the score. As of the NAIC’s November 3, 2026 map, 27 states plus DC and Puerto Rico have adopted its AI model bulletin. The bulletin covers underwriting, pricing, claims and fraud models by name. In the EU, life and health pricing models become high-risk under the AI Act from 2 December 2027.
Key Takeaways
Start from an insurer decision (quote, route, refer to SIU, retain, reserve or reinsure), then define the target and the moment of scoring. Underwriting and pricing, claims triage and fraud, lapse and retention, reserving, and catastrophe risk each need different labels, data, validation and KPIs. GLMs remain the filing default, and gradient boosting earns its place as a challenger or where its constraints and explanations can be documented. As of the NAIC’s November 3, 2026 map, 27 states plus DC and Puerto Rico have adopted the NAIC AI bulletin. Colorado, New York, California and Texas have their own rules or guidance. Proxy-discrimination testing and vendor audit rights are written regulator expectations, not optional extras. Set fraud and triage thresholds from team capacity, and measure retention models against a holdout group. Watch on YouTube
Reinventing Insurance Operations with Data and AI
Kanerika looks at how insurers rebuild underwriting, claims and service operations on data and AI, the groundwork every predictive model here depends on.
The Friday Claim That Looked Routine
Late on a Friday, a rear-end auto claim comes in. The adjuster sees a low-speed collision, a sore neck and a clean history, so the file drops into the fast-track queue.
Four months later, the same file carries an attorney letter, three treating providers and a reserve raised twice. It is now the costliest open claim in the unit, and nobody can say when it turned.
A severity model scoring at first notice would not have known the ending. Even so, the injury type and the reporting delay would have sent the file to a senior adjuster on day one.
That is predictive analytics in insurance at its most practical, a score that reaches the person who can still change the outcome. Yet the algorithm is rarely the slow part. Tying each score to a decision owner and its governing rules is where the five model families below differ.
What Predictive Analytics in Insurance Does (and Where It Stops)
Predictive analytics in insurance estimates the chance or size of a future event for a policy, a claim or a book of business. The NAIC’s AI model bulletin describes a predictive model as one that mines historic data to find patterns and predicts outcomes that support decisions. That definition is useful because it ties the model to a decision, not to a dashboard.
Descriptive reporting, the job of insurance business intelligence , shows last quarter’s loss ratio. A predictive model, by contrast, shows which new submission is likely to run hot. The wider set of insurance data analytics use cases sits around both.
In other words, every working model follows the same short chain.
Data is collected from policy, claims, billing and outside sources. Features are built from what was known at the moment of the decision. A model turns those features into a score. The score triggers a decision with an owner. Outcomes are then monitored and fed back into the next version. Still, a prediction is not a cause. A high lapse score says who is likely to leave, and it says nothing about whether a discount would keep them. The broader theory sits in Kanerika’s guide to AI predictive analytics , while this post stays with the insurer’s own models.
Five Insurance Predictive Models at a Glance: Data, KPI and Regulator Frame
Insurers run many models, but five families carry most of the money and most of the regulatory attention. The table sets out what each one predicts, what it runs on and which rule matters most.
Model Prediction target Core data Typical model Business KPI Main regulatory frame Underwriting and pricing Claim frequency, severity, loss cost Exposure, rating factors, prior losses GLM, gradient boosting as challenger Loss ratio by risk band, referral hit rate Rate filings, NY CL7, Colorado, EU Annex III 5(c) Claims severity and triage Ultimate incurred cost, litigation risk First notice details, injury codes, notes Gradient boosting, survival or text models Severity leakage, cycle time, escalation accuracy NAIC bulletin, unfair claims practice laws Fraud and SIU scoring Likelihood a claim merits investigation Claim attributes, provider and party links, history Supervised classifier, anomaly and network models SIU hit rate, net recoveries NAIC bulletin, claims handling rules Lapse and retention Lapse or non-renewal within a horizon Billing, tenure, premium change, claims and service contacts Classifier or survival model Incremental retention against a holdout NAIC bulletin, rate rules when a discount is offered Reserving and catastrophe Claim development, tail loss by location Loss triangles, claim transactions, hazard and exposure data Claim-level ML as challenger, vendor cat models Reserve accuracy over time, probable maximum loss Actuarial standards, NAIC bulletin third-party section
Notice the KPI column. A model can rank risks well (a strong AUC) and still fail its KPI if the score arrives after the adjuster has already decided.
Technical metrics prove the model ranks well, and business KPIs prove someone used it. The algorithm side, meanwhile, is covered under machine learning in predictive analytics .
The sections that follow take the five families one at a time.
Underwriting and Pricing Models: GLMs, Gradient Boosting and the Rate Filing
Underwriting and pricing models decide whether to write a risk and at what price. They are the oldest predictive models in insurance, and they draw the closest scrutiny because their output reaches the consumer as a premium.
Risk Selection at Quote
A risk selection model predicts expected loss for a submission over a defined exposure period. Every input must therefore be available at the moment of quote, so a feature built from the first year of claims cannot appear in it.
In practice, most carriers use these scores to refer rather than decide. Typically, a high score sends the file to an underwriter with the reasons attached, while low-risk submissions flow straight through. The KPIs that matter are the referral hit rate and the loss ratio by risk band a year later.
Frequency, Severity and Pure Premium
Pricing actuaries usually model claim frequency and claim severity separately, because the two respond to different drivers, then combine them into a pure premium. Meanwhile, exposure offsets keep a six-month policy from looking safer than a twelve-month one. A Tweedie model can fit loss cost directly when the team wants one model instead of two.
Even so, the predicted loss cost is not the filed premium. Expenses, profit loading, caps on rate change and the state’s approved rating plan all sit between the model and the price a customer sees.
GLM vs Gradient Boosting for a Rate Filing
Generalized linear models remain the default for filed rating plans. The NAIC’s Regulatory Review of Predictive Models white paper was adopted by its property and casualty committee in December 2020. It gives state reviewers a detailed list of information to request for plans built wholly or partly on a GLM. Gradient boosting often finds interactions a GLM misses, yet it must answer the same questions.
Factor GLM Gradient boosting Interpretability Each factor has a visible relativity Needs SHAP values or partial dependence plots Interactions Only those the actuary specifies Found automatically, which can hide proxies Monotonic behavior Set by the model form Must be added as explicit constraints Filing documentation Familiar to most state reviewers Longer, with more questions expected
Therefore, a sound pattern is to run gradient boosting as a challenger on an out-of-time cohort. Where it beats the GLM, the actuary can pull the interaction it found back into the GLM as a documented factor.
Regulator Frame for Pricing Models
First, the NAIC bulletin reminds insurers that rates must not be excessive, inadequate or unfairly discriminatory, whatever method produced them. New York’s Circular Letter No. 7 applies specifically to AI systems and external data in underwriting and pricing. It expects testing for disproportionate adverse effects before production.
Similarly, Colorado’s amended Regulation 10-1-1 sets governance duties for life, private passenger auto and health insurers that use external consumer data and predictive models. In the EU, pricing for life and health insurance is a high-risk use under the AI Act, with obligations from 2 December 2027. Kanerika’s guide to explainable AI also explains the reason codes these frames expect.
Claims Triage, Severity and Fraud Scoring
Unlike pricing models, claims models act on money that is already moving, which makes them the fastest place to show value. They also carry the hardest label problems in the business, because the true outcome of a claim can take years to settle.
Predicting Ultimate Severity at First Notice
First, a severity model estimates what a claim will finally cost. The score at first notice of loss (FNOL) is the one that matters, since it decides who handles the file. A second score after the first review adds notes, medical codes and adjuster actions, and it is usually more accurate but arrives later.
Open claims, however, are the trap here. A model trained only on closed claims learns from the files that settled fast.
Those are the cheap ones, so the model underestimates the slow and expensive tail. Survival methods or development-adjusted targets, by contrast, handle that censoring better.
Accordingly, the KPIs are severity calibration by band, escalation accuracy and leakage, meaning money paid that tighter handling would have avoided. The same pipeline can score subrogation potential, so recoverable claims reach the recovery team before evidence goes stale.
Fraud Scoring for SIU Referral
In its 2022 study , the Coalition Against Insurance Fraud estimated that fraud costs American consumers at least $308.6 billion every year. It also puts fraud at about 10% of property and casualty losses. Those numbers explain why fraud scoring is often the first claims model an insurer funds.
Good fraud programs blend three signal types. Supervised classifiers learn from confirmed cases, and anomaly detection flags claims that look unlike their peers.
Network analysis, for example, finds shared phones, addresses or repair shops across parties. Kanerika’s guide to AI in fraud detection compares these methods across industries.
Fraud labels are also delayed and selective. Only referred claims get investigated, so the training data reflects past referral habits as much as real fraud.
So set the threshold to what the special investigations unit (SIU) can work each week. Then measure the hit rate and net recoveries within that capacity.
Above all, a referral is not a fraud finding. The score buys an investigator’s attention, and the decision to deny or pursue still belongs to a person with documented reasons.
Watch on YouTube
Kanerika | Revolutionizing Fraud Detection in Insurance with AI/ML-Powered RPA | AI/ML | RPA
A Kanerika walkthrough of insurance fraud detection that pairs machine learning scores with automated claim handling, so investigators see the riskiest files first.
Regulator Frame for Claims and Fraud Models
The NAIC bulletin names claim management and fraud detection within its scope, so these models belong in the insurer’s written AI program. Unfair claims settlement practice laws still apply to any decision a score influences, which is why human review and reason codes matter.
In the EU, claims triage and fraud scoring are not named in Annex III point 5(c) of the AI Act. EIOPA’s August 2025 opinion still clarifies how existing insurance law applies to these AI systems. Document assistants that summarize claim files are a separate topic, covered under generative AI in insurance .
Lapse, Retention and Renewal Models
Retention models start with a definition problem. Lapse for non-payment, a mid-term cancellation and a non-renewal are three different events with different causes. A model that blends them will therefore predict none of them well.
The usual inputs are billing behavior, tenure, the premium change at renewal, recent claims experience and service contacts. A classifier, for example, answers whether a policy will lapse in the next 90 days. A survival model also answers when, which helps life and annuity teams plan outreach.
That is why the KPI to report is incremental retention against a randomized holdout, not how well the model ranks risk. Many high-risk policyholders would leave whatever the insurer does, and many low-risk ones would stay anyway. The customer churn prediction guide covers model building in detail, and customer retention analytics covers what happens after the score.
Retention offers also carry a regulatory edge. A renewal discount is a pricing decision, so rate rules and fairness testing apply to it. Campaign execution for cross-sell sits closer to predictive analytics in marketing .
AI Assessment
Is Your Insurance Data Ready for Predictive Models?
Score your data, model and governance foundations in minutes and see which insurance model, from pricing to claims triage, is a realistic starting point.
Start Your AI Assessment → Reserving and Catastrophe Risk Scoring
Reserving and catastrophe models work on the balance sheet rather than on single customers. Their errors show up later and larger, so they need the most careful backtesting.
Claim-Level Reserving as a Challenger to Chain-Ladder
Traditionally, chain-ladder methods project aggregate paid and incurred triangles forward. Claim-level machine learning models predict development for each open claim from its transactions, notes and inflation exposure, then roll the results up.
The test is a rolling backtest on past valuation dates, using only data available on each date. Passed that way, these models support actuarial judgment and help explain movements.
They do not, however, replace the appointed actuary’s sign-off. Kanerika’s guide to AI in risk management covers similar backtesting habits. By contrast, predictive analytics in banking works on credit-risk horizons rather than loss development.
Catastrophe and Property Risk Scoring
Catastrophe scoring combines hazard, exposure, vulnerability and financial terms such as limits and deductibles. In particular, geospatial features like distance to coast, wildfire fuel and elevation drive much of the signal at the property level.
Because reinsurance and accumulation limits are set on the worst years, tail calibration matters more than average accuracy here. In addition, most insurers license vendor catastrophe models with machine learning components inside, and those fall under the bulletin’s third-party section. The scores feed accumulation limits, underwriting appetite and reinsurance buying.
Predictive Analytics in Insurance by Line: Life, Health, P&C and Workers’ Comp
The five model families look different in each line of business. The data changes, and so does the rule a model must satisfy.
Line Main predictive uses Regulator note Life Accelerated underwriting, mortality, lapse Colorado 10-1-1, NY Circular Letter 7, EU Annex III 5(c) Health Risk adjustment, utilization, high-cost claimants Colorado health benefit plan rules, EU Annex III 5(c) P&C and auto Pricing, telematics, claims severity, fraud Rate filings, Colorado private passenger auto rules Workers’ comp Claim duration, return to work, litigation risk State claims handling rules, NAIC bulletin
Overall, life and health carry the heaviest stack of AI-specific rules as of November 2026. P&C pricing, meanwhile, still runs through state rate filing review.
Checklist
AI Governance Checklist
A practical checklist for model inventory, ownership, validation and monitoring, so each pricing, claims or fraud model has the evidence a regulator will ask for.
Get the Checklist → The Regulator Map: Which Rules Apply to Which Model
Insurance AI rules come from three layers, the NAIC model bulletin, a handful of state rules that go further and the EU AI Act. Each layer, in turn, reaches a different set of models.
The NAIC AI Model Bulletin and State Adoption
To begin with, the NAIC adopted its bulletin on the use of AI systems by insurers on December 4, 2023. Because it is guidance, it binds an insurer only where a state adopts it. As of the NAIC’s November 3, 2026 map , 27 states plus DC and Puerto Rico have adopted it.
In addition, the bulletin expects a written AI systems program covering models the insurer builds and those it buys. The program may draw on a framework such as the NIST AI Risk Management Framework. Kanerika’s AI governance framework guide shows how such programs are usually structured.
State Rules That Go Further
The same NAIC map lists four states with their own insurance AI rules or guidance instead.
Colorado’s SB21-169 and amended Regulation 10-1-1, effective October 15, 2025, set governance duties for life, auto and health insurers. Its quantitative testing rule for life underwriting is still a draft. New York’s Circular Letter No. 7, issued July 11, 2024, covers AI and external data in underwriting and pricing. California’s Bulletin 2022-5 dates from June 30, 2022. Texas Bulletin B-0036-20 dates from September 30, 2020. The EU AI Act for Life and Health Pricing
The EU AI Act treats some insurance AI as high-risk. Annex III point 5(c), in particular, covers risk assessment and pricing of natural persons in life and health insurance.
Regulation (EU) 2026/1744 moved those obligations to 2 December 2027. Kanerika’s post on EU AI Act compliance also covers the wider duties.
Model type NAIC bulletin Colorado and New York EU AI Act Evidence to keep Underwriting and pricing In scope NY CL7 applies; Colorado for life, auto and health High-risk for life and health from 2 Dec 2027 Filing support, fairness tests, data sources Claims triage In scope Colorado governance for covered lines Not named in 5(c) Routing rules, human review records Fraud and SIU In scope Colorado governance for covered lines Not named in 5(c) Referral reasons, outcome logs Lapse and retention In scope Pricing rules if offers change premium Not named in 5(c) Offer rules, holdout results Reserving and catastrophe In the AI program; vendor section applies Limited direct reach Not named in 5(c) Backtests, vendor documentation
This map reflects the rules as reviewed in November 2026. Recheck the NAIC map before any filing, since states keep joining it.
Fairness and Proxy-Discrimination Testing, Step by Step
Removing race, religion and other protected attributes from a model does not, by itself, make it fair. ZIP code, credit-based insurance scores, occupation and even vehicle type can stand in for them. That is why Colorado’s SB21-169 and New York’s Circular Letter No. 7 both ask insurers to test outcomes, not just inputs.
A workable testing loop has six steps.
Inventory every input, including external consumer data and vendor scores. Test whether each input correlates with protected classes. Run outcome tests such as the adverse impact ratio, standardized mean differences and denial odds ratios. Trace the drivers of any disparity back to specific variables. Record whether a less discriminatory alternative exists and why the chosen design is needed. Repeat the tests after every material change to data or model. New York, for instance, lists these metrics, along with z-tests and t-tests, as examples and encourages insurers to use more than one. However, the circular also says insurers are not expected to collect more personal data to run the tests. Inference methods can estimate group membership instead, and Kanerika’s guide to responsible AI covers the wider fairness auditing these tests belong to.
Third-Party Model and Data Vendor Oversight
Most insurers buy part of their predictive stack, such as credit-based insurance scores, catastrophe models, fraud consortium data and telematics scores.
Even so, Section 4.0 of the NAIC bulletin keeps the insurer accountable for those systems. New York goes further and says an insurer may not rely solely on a vendor’s claim that its data or model is non-discriminatory. A practical vendor file holds six things.
Due diligence notes on how the vendor built and validated the model. Documentation of training data sources and known limitations. Audit rights, or the right to audit reports from qualified auditors, where available. A clause requiring the vendor to cooperate with regulatory inquiries. Notice terms for any material model or data change. An exit plan and fallback if the score is withdrawn. Procurement and legal own this list, yet only data scientists can judge a vendor’s validation report. Kanerika’s guide to machine learning governance also covers model inventories that track vendor models next to internal ones.
Data Foundation and the Path from Pilot to Production
Insurance models fail on data far more often than on algorithms. Policy, exposure, claims and billing data usually live in separate systems, each with its own history of codes and corrections.
Point-in-Time Data and Label Maturity
Above all, every feature must reflect what was known on the scoring date. A severity model that sees later payments, or a fraud model that sees SIU outcomes, will look brilliant in testing and fail in production.
Labels also need time to mature, because claims develop slowly. Long-tail liability claims can stay open for years, so the team must decide how far back to train and how to treat open files.
Telematics and third-party data also need provenance records, and data lineage tooling makes that traceable. A data quality framework catches broken feeds before they reach a model.
Kanerika’s post on Databricks insurance analytics walks through that pattern, and a Databricks feature store keeps training and scoring features consistent.
Datasheet
Insurance Data Ingestion Accelerator for Databricks Lakehouse
How Kanerika brings policy, claims and billing data into a Databricks lakehouse, the point-in-time foundation that insurance pricing, claims and fraud models train on.
Download the Datasheet → Deploying and Monitoring Scores Inside Core Systems
Batch scoring suits renewal lists and reserve reviews. Event scoring fits first notice of loss, while an API call suits quote flows.
Before go-live, a new model should run in shadow mode, scoring live cases without acting on them. The platform should also log the model version, inputs and reason codes for every score.
Once the model is live, monitor drift, calibration and subgroup outcomes, and send any retrained version through the same approval gate. Kanerika’s guide to machine learning operations covers the tooling.
What Insurers Gain, and Where Projects Stall
The gains are concrete, but only when a model is wired to a decision. Underwriters see sharper risk selection, and complex claims reach senior adjusters sooner. SIU investigators spend time on better leads, retention teams act before renewal, and actuaries get tighter reserve ranges to debate.
Fraud shows the scale: with fraud in about 10% of property and casualty losses, a sharper SIU referral list moves real money.
Still, projects stall in predictable places. Labels arrive late, legacy policy and claims systems resist real-time scoring, and handoffs between actuaries and data scientists blur ownership. Vendor black boxes slow approvals, and governance added after the build forces rework, so agree the decision, label and evidence pack first.
Talk to Kanerika
Plan Your First Insurance Model With Kanerika
Bring one decision, such as claims triage or SIU referrals, to Kanerika’s team and map the data, owner and regulator evidence before you build.
Book a Meeting → How Kanerika Helps Insurers Put Predictive Models into Production
Kanerika’s predictive analytics services page makes one delivery point: plan for deployment from day one. Its teams run a rapid validation phase on real data before committing to a full-scale build. A well-scoped pilot, per the page, runs 6 to 10 weeks from data assessment to initial production deployment.
Multi-model engagements with heavy data engineering take 3 to 6 months. For an insurer, that gap argues for starting with one decision and one owner, such as the SIU referral queue. The work also keeps governance audit trails, and in insurance that trail becomes the evidence pack a regulator asks for.
Kanerika’s insurance industry page lists predictive models that improve underwriting accuracy, detect fraud and improve claim outcomes. Model engineering runs through its machine learning services , and reporting through its data analytics services . Kanerika holds ISO 27001, ISO 27701 and SOC 2 Type II certifications.
One client was a mid-sized US insurer rated A+ by AM Best. Kanerika built fraud detection with Isolation Forest and autoencoder models and added financial risk assessment and forecasting models. The data moved onto Azure Data Lake behind multi-factor authentication, and Microsoft tooling cut the manual effort in data transfers and EXIM report generation.
As a result, the insurer recorded a 33% decrease in financial losses and a 25% reduction in overall risk exposure.
Case Study
33% Fewer Financial Fraud Losses with AI Modeling & Forecasting
For a US insurer rated A+ by AM Best, Kanerika built fraud detection and risk models and moved the data onto Azure Data Lake. Financial losses fell 33% and risk exposure 25%.
Read the Case Study → Start With One Insurance Decision and Its Rulebook
Predictive analytics in insurance works when each model answers one decision. That decision might be a quote, a claim route, an SIU referral, a retention offer or a reserve review. So define the target and the moment of scoring, then match the regulator frame before anyone trains a model.
Finally, three questions show whether a team is ready. Who acts on the score, and how many cases can they handle? Can the data be rebuilt as it looked on the scoring date? Which rule applies, and what evidence will the reviewer ask to see?
Frequently Asked Questions
What is predictive analytics in insurance?
Predictive analytics in insurance uses policy, claims, billing and outside data to estimate what is likely to happen next for a policy or claim. Typical outputs are an expected loss cost, a claim severity score, a fraud referral score or a lapse probability. Each score is useful only when it feeds a named decision.
How do actuaries use predictive analytics in insurance?
Actuaries use predictive models to set rates, estimate reserves and test assumptions. In pricing they often model claim frequency and severity separately with GLMs. In reserving they may run claim-level machine learning as a challenger to chain-ladder methods, backtested on past valuation dates, while keeping their own judgment and sign-off.
How is predictive analytics used in insurance underwriting?
Underwriting models predict the expected loss for a new submission using only data available at quote. Most carriers use the score to refer risky files to an underwriter with reasons attached, while clean risks flow straight through. Success is measured by referral hit rate and by each risk band’s later loss ratio.
How do insurers use predictive analytics for claims and fraud detection?
Claims models estimate ultimate severity at first notice of loss, so complex files reach senior adjusters early. Fraud models blend supervised classifiers, anomaly detection and network analysis to rank claims for SIU review. The referral threshold should match what investigators can work each week, and a referral is never treated as a fraud finding without human review.
What is the difference between a GLM and machine learning in insurance pricing?
A GLM shows a clear relativity for every rating factor, which state reviewers know well, so it remains the filing default. Gradient boosting and similar machine learning models find interactions automatically and often predict better, but they need extra explanation tools and monotonic constraints. Many insurers run them as challengers and move useful findings back into the GLM.
Which states have adopted the NAIC AI model bulletin?
As of the NAIC’s November 3, 2026 implementation map, 27 states plus the District of Columbia and Puerto Rico have adopted the bulletin. California, Colorado, New York and Texas have their own insurance AI rules or guidance instead. States keep joining, so insurers should check the current NAIC map before filing or launching a new model.
Is insurance pricing high-risk under the EU AI Act?
Only part of it. Annex III point 5(c) of the EU AI Act lists AI used for risk assessment and pricing of natural persons in life and health insurance as high-risk. Regulation (EU) 2026/1744 set those obligations to apply from 2 December 2027. Property and casualty pricing, claims and fraud models are not named in that point.
How do insurers test predictive models for proxy discrimination?
They inventory every input, test whether inputs correlate with protected classes, and run outcome tests such as the adverse impact ratio and standardized mean differences. Then they trace any disparity to its drivers, document less discriminatory alternatives and retest after each material change. New York does not expect extra personal data collection.
How is predictive analytics different from generative AI in insurance?
Predictive analytics produces a score or estimate, such as a loss cost, a fraud referral score or a lapse probability, that feeds a decision. Generative AI produces text or summaries, for example drafting claim notes or answering policy questions. Both fall under the NAIC bulletin when they support regulated decisions, but they are built, tested and monitored differently.
