Industries / Finance and business services / Insurance
Pricing and underwriting
Deciding whether to accept a risk, on what terms, and at what price; the deciding is called underwriting.
7
Milestones achieved, from 1693 to 2022.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1693
Deaths counted by age
Edmond Halley used records of births and funerals from the city of Breslaw to estimate how many people die at each age, in a paper whose title ends "with an Attempt to Ascertain the Price of Annuities upon Lives". An annuity is a payment made for as long as a person lives. The St Andrews MacTutor archive calls his tables one of the earliest works to relate mortality and age, and "highly influential in the future production of actuarial tables in life insurance".
Credit
Edmond Halley, Philosophical Transactions of the Royal Society, volume 17
1762
Premiums set by age
The Equitable Society in London sold life insurance with premiums that stayed level for the life of the policy and were graded by the age of the person insured. The early overpayments went into a fund for the higher death rates of later years. The method was devised by the mathematician James Dodson.
Credit
James Dodson; the Society for Equitable Assurances, later Equitable Life
2000
A free guarantee comes due
From 1957 to 1988 Equitable Life sold pension policies that guaranteed to turn each £100 saved into an income of about £10 a year at retirement, and charged nothing extra for the guarantee. When interest rates fell, the guarantee became valuable. On 20 July 2000 the House of Lords ruled that the Society could not pay those policyholders smaller bonuses to cover it, which added £1.5 billion to its liabilities, and it closed to new business that December.
LessonA guarantee sold for nothing can turn into an insurer's largest debt when markets move.
Credit
Dalit Baranoff and Christopher O'Brien, whose 2016 case study for the Geneva Association sets out the sequence
2007
Credit scores put to test
Many US car insurers price with credit-based insurance scores, numbers built from a person's credit record. The Federal Trade Commission found the scores "effective predictors of risk", and found that they are distributed differently among racial and ethnic groups. It also found that little of their predictive power comes from that link, and it could not build a score that predicted as well with smaller gaps between groups.
LessonWhether a factor predicts claims and whether it falls evenly on groups are separate questions, and each needs its own test.
Credit
US Federal Trade Commission, report to Congress, July 2007
2008
Priced by how you drive
Progressive piloted a program called TripSense in 2004 to study driving habits, launched MyRate in 2008 to give "a customized rate based on how they actually drive", and renamed it Snapshot in 2011. The NAIC, the association of US state insurance regulators, calls such usage-based insurance "increasingly more common" and has asked insurers about it in a yearly data call since 2020.
Credit
Progressive
2011
One price for both sexes
On 1 March 2011 the Court of Justice of the European Union ruled that using a person's sex as a risk factor in insurance is discrimination. From 21 December 2012, insurers in the EU had to charge women and men the same premiums and pay them the same benefits.
Credit
Test-Achats, a Belgian consumer association, which brought the case; the Court of Justice of the European Union
2022
Loyalty penalty banned
UK insurers had charged customers who renewed more than new customers for the same cover, a practice called price walking. From 1 January 2022 the Financial Conduct Authority (FCA) required home and motor insurers to offer a renewal price no higher than the new-customer price. The FCA estimated savings for consumers of £4.2 billion over 10 years.
Credit
UK Financial Conduct Authority
Watch for
Quotes from email in seconds
CFC, a specialty insurer, says its pilot system Lane Assist is designed to take an emailed cyber insurance application to a quote recommendation "in seconds". It handles "a limited number" of real new applications, and an underwriter approves every quote before it is issued.
Achieved whenan outside party reports the system's speed and its error rate against human underwriters on a stated number of applications.
Claimed by
CFC, April 2026
Open
01
Car-pricing factors in public
A free table that shows, for the ten largest car insurers in five US states, how much each filed rating factor moves the premium for one standard driver. Factors include age, credit-based insurance score, years licensed, home ZIP code and past claims. Each number links to the page of the rate filing it came from.
The briefCredit · unclaimed02
Testing prices for bias
Open-source code that runs the tests in Colorado's 2023 draft rule for life insurers. It estimates applicants' race and ethnicity from first name, surname and address with BIFSG (Bayesian Improved First Name Surname Geocoding). It then compares decline rates and premiums per $1,000 of cover between groups and flags gaps of 5% or more. It ships with made-up applicants that carry gaps of known size, so anyone can check that the code finds them.
The briefCredit · unclaimed
Open 01 · the brief
Car-pricing factors in public
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free table that shows, for the ten largest car insurers in five US states, how much each filed rating factor moves the premium for one standard driver. Factors include age, credit-based insurance score, years licensed, home ZIP code and past claims. Each number links to the page of the rate filing it came from.
Why it is still open
Rate filings are public through state portals and SERFF, the regulators' shared filing system. The one open-source effort found, QuoteFii's Auto Insurance Data Tools, covers 3 of 51 jurisdictions and works from regulators' sample-premium tables rather than from the filings.
Built with what exists
state insurance department filing portals and SERFF Filing Access, downloaded by hand within each site's terms; a language model such as Claude or Gemini to read rate manuals and fill the table, citing the page for each value (see the Note "Seven more flagged"); Python and pandas to compute how much each factor changes the price.
First step
Download the current private car insurance rate filing of the largest insurer in one state and list its rating factors by hand, as the answer key for the model.
Done looks like
50 insurer-and-state pairs, with every factor traced to a filing page. A second person checks 50 values picked at random and agrees with at least 48. The method is published beside the table.
Open 02 · the brief
Testing prices for bias
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Open-source code that runs the tests in Colorado's 2023 draft rule for life insurers. It estimates applicants' race and ethnicity from first name, surname and address with BIFSG (Bayesian Improved First Name Surname Geocoding). It then compares decline rates and premiums per $1,000 of cover between groups and flags gaps of 5% or more. It ships with made-up applicants that carry gaps of known size, so anyone can check that the code finds them.
Why it is still open
Colorado's 2021 law, SB21-169, requires insurers that use outside consumer data and models to test them for unfair discrimination, and the 2023 draft set out how. In October 2025 the Colorado Division of Insurance cited "the absence of finalized rules" and waived the testing descriptions in life insurers' annual reports for 2024 and 2025.
Built with what exists
Python; an open-source race-estimation package such as surgeo (Python) or wru (R); US Census surname and geography tables; a script that generates made-up applicants.
First step
Write the draft rule's two tests as a one-page specification and generate 10,000 made-up applicants with one planted gap.
Done looks like
On 100,000 made-up applicants with planted gaps of 2%, 5% and 10%, the code flags the 5% and 10% gaps in at least 95 of 100 runs, and the 2% gap in no more than 5 of 100. Code and data are published under an open license.