Solutions · Insurance

Pricing models you can defend in a filing.

Open up the GLMs and gradient-boosted models behind rating, underwriting and claims. See what each rating factor does, explain any quote in its own units, and know when a new channel changes the book.

Data
678,013French motor liability policies, 358,360 policy-years of exposure
Lift
9×between the lowest and highest predicted deciles: 3.9 to 35.5 claims per 100 policy-years
A factor with no weight
0.0%of importance for area band, once town density is in the model
A new channel
PSI 7.89on driver age; expected frequency of the batch rises from 11.7 to 21.4

Where it fits

From the rating engine to the claims desk.

For the GLMs and machine-learning models behind pricing, underwriting and claims.

Pricing

Frequency and severity models

GLMs and gradient-boosted models for rating. See what each factor does across the book, and whether it does it in one direction.

Underwriting

Risk selection and triage

Why this risk was referred, from the model's own inputs, stored with the quote.

Claims

Triage and fraud flags

Explain the score behind a fast-track or a referral, so the handler sees what the model saw.

Filing

Support for rate filings and reviews

Importance, behaviour curves and learned rules in plain English, exportable as JSON, with the seed that reproduces them.

Portfolio

Mix and channel monitoring

Know within one batch when a new broker, product or region changes who is being quoted.

Third parties

Vendor and legacy models

A model in another language, or one you may not open, is explained through its scoring endpoint or from attributions you compute.

A worked example

A motor frequency model, opened up.

Public datafreMTPL2freq, CASdatasets, via OpenML (Dutang & Charpentier).motor-frequencyscikit-learn · gradient boosting

A Poisson gradient-boosted model predicts third-party claim frequency from seven rating factors, weighted by exposure and trained on 542,410 policies. Kaimei explains each quote in the model's own unit, here claims per 100 policy-years.

motor-frequency · Explain
scikit-learn · gradient boosting
raiseslowers
freMTPL2freq POST https://api.kaimei.ai/v1/explanations

Does it rank risk?

Nine times the claims, bottom to top.

On 135,603 policies the model never saw, observed claim frequency climbs from 3.9 to 35.5 per 100 policy-years across predicted deciles, and predicted frequency tracks it.

Policy A is in that top decile at 45.1. The explanation splits the 33.7 above average into its parts: +17.9 for an entry-level bonus-malus of 100, +10.0 for a driver aged 20, and small amounts for the rest.

Observed claims by predicted decile135,603 held-out policies
41
2
3
4
5
6
7
8
9
3610
Claims per 100 policy-years, lowest to highest predicted decile: 3.9 observed against 4.3 predicted in the first, 35.5 against 33.9 in the tenth. The model explains 7.1% of Poisson deviance.

Rating factors

Each factor's effect, and its direction.

A behaviour curve varies one factor and holds the others as they are. Bonus-malus is monotonic: Higher bonus_malus increases claim_frequency, most sharply between 80 and 100 (range 0.0779 → 0.243).

Driver age is not, and Kaimei says so: driver_age has a non-monotonic effect on claim_frequency: it bottoms out around 30 years (range 0.0815 → 0.133).

And one factor does nothing. Area band carries 0.0% of measured importance, because town density already holds the information. That is worth knowing before it goes in a filing.

Behaviour curve · bonus-malusincreasing · consistency 1.00
50
60
80
Average predicted claims per 100 policy-years as bonus-malus alone is varied across the reference policies. It rises at every step.
Global importance · motor-frequencypermutation · 2,000 rows
Vehicle age29.0%
Bonus-malus28.5%
Driver age15.6%
Power band12.0%
Fuel8.3%
Town density6.6%
Area band0.0%
vehicle_age is the most influential input, followed by bonus_malus and driver_age. Together they carry 73% of the measured importance (permutation importance over 2000 rows).

Portfolio monitoring

A new channel changes the book.

Five hundred held-out policies are logged as normal business, and every factor is stable. Then a batch arrives from a new distribution channel: drivers of 26 and under in the two densest area bands, drawn from the same data. Four factors move at once, and the batch's expected frequency is 21.4 claims per 100 policy-years against 11.7. The model is not wrong. The mix has changed, and pricing should hear about it today.

Normal businessstable
Vehicle age0.03
Area band0.02
Town density0.02
Driver age0.01
Power band0.01
Bonus-malus0.00
Fuel0.00
Recent inputs match the reference population (max PSI 0.03). 500 logged predictions against 5,000 reference rows.
The new channelsignificant
Driver age7.89
Area band5.96
Town density5.14
Bonus-malus3.01
Power band0.09
Fuel0.06
Vehicle age0.04
Recent traffic has significant drift (PSI 7.89). Shifted inputs: driver_age, area, density, bonus_malus. 500 logged predictions against 5,000 reference rows.

Population stability index per input. Bars are cut off at 0.5; the line marks 0.25, where Kaimei calls drift significant.

Who it is for

Actuaries, governance and the front line.

01

Pricing actuaries

See each factor's effect and direction on the fitted model, find the factors that carry no weight, and compare a new rating model with the one it replaces.

02

Model governance and compliance

Keep an explanation with every quote and referral, and export a model's insights for the governance file with the seed that reproduces them.

03

Underwriting and claims

Read why a risk was scored the way it was, in the dashboard, in the model's own unit, without opening a notebook.

Rules and guidance

Evidence for model governance and filings.

Rule or guidanceWhat it asksWhere Kaimei helps
NAIC Model Bulletin on the Use of AI Systems by InsurersAdopted December 2023; in place in 28 states, the District of Columbia and Puerto Rico as of November 2026A written programme for governance, risk management and controls over AI systems, including predictive models.Documentation of what each model relies on, validation evidence, and monitoring for drift.
Colorado Regulation 10-1-1Life, private passenger auto and health benefit plan insurersA governance and risk management framework for external data, algorithms and predictive models.Model documentation, stored explanations and scheduled monitoring.
New York DFS Insurance Circular Letter No. 7 (2024)Governance and transparency for AI systems and external consumer data in underwriting and pricing.The inputs behind each decision, on the record, to draw a consumer explanation from.
EU AI ActAnnex III: risk assessment and pricing for life and health insurance. High-risk obligations apply from 2 December 2027Risk management, logging, transparency, human oversight, accuracy and robustness.Logged predictions and explanations, robustness and drift analyses.
Solvency II internal modelsDirective 2009/138/EC, Articles 120 to 126A use test, statistical quality standards, validation and documentation.Sensitivity, stability and behaviour evidence for the validation report.

Coming soon

Next on the roadmap.

What we are building now, and what to use until each one ships.

  • Coming soonUnfair discrimination and proxy testing

    Outcome rates compared across the groups you define, and a check for inputs that stand in for a protected characteristic. Until then, explanations help investigate a concern; they are not a test.

  • Coming soonCategorical tree splits, run directly

    LightGBM, CatBoost and XGBoost models with categorical splits read and run by Kaimei itself. Until then, they are explained through a scoring endpoint.

  • Coming soonFrequency and severity in one model

    Several outputs per model, so frequency, severity and the pure premium are explained in one call. Until then, register them as separate models.

  • Coming soonFiling-ready documentation

    Rating factors, behaviour curves and learned rules exported as a document for a rate filing or a model review. Until then, they export as JSON.

See it on your own model.

Bring a rating model. We will show you what each factor is doing.