intelligentactuaries/synthetic-lapse-surrender-study-2023-2025

Synthetic Term & Whole Life Lapse and Surrender Study, 2023–2025

10,000 synthetic policyholders · 27,692 VM-51 policy-year records · 25,776 policy-years of exposure · ten markets

This is synthetic data. Every record was generated, not extracted. It is a documented, reproducible generative prior about policyholder behaviour, not experience. It carries no information about any real portfolio or person, and no valuation, pricing or reserving assumption should be set from it. Please label it as synthetic wherever it goes.

This dataset answers the Society of Actuaries (SOA) Research Institute and LIMRA 2023–2025 Term and Whole Life Lapse and Surrender Experience Study data request (July 2026) from the other side. The request asks life insurers for VM-51 policy-level records of their individual term and whole life business for calendar years 2023 to 2025. Intelligent Actuaries carries no such book, so we generated one. Each of 10,000 synthetic policyholders (an agent in the Scelo swarm's society tier, running on Claude Code with claude-sonnet-5) describes the one individual term or whole life policy they hold and what happened to it in 2023, 2024 and 2025. A deterministic driver validates every answer against the product contract, overlays an independent mortality decrement, and expands each history into VM-51 policy-year records in the Appendix B layout the request specifies.

🚀 Explore it live: Lapse & Surrender Study · scelo, a Space built on the scelo Python library: experience tables, the lapse GLM, and lapse-adjusted pricing, IFRS 17 CSM and Solvency II life SCR on the term book, recomputed as you change the basis.

Acknowledgements

We gratefully acknowledge the Society of Actuaries Research Institute, the research arm of the Society of Actuaries (SOA), and LIMRA, who run the Term and Whole Life Lapse and Surrender Experience Study together. This dataset would not exist without their work.

• Their 2023–2025 data request (July 2026) defines everything this dataset imitates: the scope (individual level-premium term, annually renewable term, and participating and non-participating whole life), the study window (calendar years 2023 to 2025), and the policy-level record layout (the VM-51 format, reproduced as the request's Appendix B).
• Their 2015–2022 Term and Whole Life Lapse and Surrender Experience Study (2024), the edition this request updates, is the published reference for how this experience is studied.
• The record format itself is the NAIC Valuation Manual's VM-51 (Experience Reporting Formats), which the request adopts for policy-level data.

Independence. This is independent work by Intelligent Actuaries. It is not an SOA or LIMRA publication or dataset, and it is not affiliated with, sponsored by or endorsed by the SOA, the SOA Research Institute, LIMRA or the NAIC. It contains no SOA, LIMRA or contributing-company experience. The data-request document is © 2026 Society of Actuaries Research Institute and LIMRA and is not redistributed here; read it on the SOA website. Any errors are ours.

What is in the repository

data/        the tabular files, one viewer config each (table below)
vm51/        the submission-format file: fixed-width ASCII, Appendix B
raw/         every model call, raw reply and parsed policy (JSON Lines)
metadata/    design manifest, the dated per-market context, the failures log
analysis/    the term-book analysis on three engines, and every figure the study quotes

Config	File	Rows	One row per
vm51_policy_years (default)	data/vm51_policy_years.csv	27,692	policy × observation year: VM-51 Appendix B items 1–32 and 47 (amounts in USD), plus study extensions (local-currency amounts, market, MBTI type, demographics, central exposure, the policyholder's note for the year)
policies	data/policies.csv	10,000	policy: product, amounts, status in each of 2023, 2024 and 2025, termination timing, notes, voice, repair log
agents	data/agents.csv	10,000	synthetic policyholder: the design (market, age, sex, MBTI type, locale, community, income, education, employment, family, mortgage, risk tolerance, financial literacy)
term_book	data/term_book_2023_2025.csv	4,254	term policy: the analysis population for project A (all term policies, with exposure in 2023–25 and final status)
term_book_inforce	data/term_book_inforce_2025.csv	2,698	term policy in force and inside its level term at 31 December 2025: project B

File	What
vm51/vm51_2023_2025.txt	The 27,692 records as the request asks for them: fixed-width ASCII, 285 characters per record, Appendix B items 1–47, with the UL/VUL items 33–46 blank
raw/responses.jsonl	13,247 lines, one per model call: the raw reply and the parsed policy (the full audit trail). 10,000 succeeded; the other 3,247 are failed attempts that were retried (3,246 Claude Code errors, during which the run paused 919 times for plan rate limits, and one unparseable reply). Nested JSON; read it with pandas.read_json(..., lines=True)
metadata/manifest.json	Design, seed, FX table, counts, status tallies
metadata/iaai-canon-country-context.json	The 30 dated context entries, three per market (economy and household finances; politics, society and security; life-insurance norms), August 2023 to August 2026. Each agent saw only its own market's block
metadata/failures.jsonl	Agents whose reply could not be parsed after a retry. Empty: all 10,000 answered
analysis/facts.json	Every figure quoted in the study's reports, computed from the files above
analysis/scelo-ide/	Projects A and B run in Scelo IDE 0.1.6: results, GLM coefficients, the monthly projection, board packs (PDF), the assistant's narratives, exported workflows (Python, R, C++, prompt), .sce project files and session snapshots
analysis/scelo-libraries/	The same analysis as code with scelo for Python and scelo for R: results, GLM coefficients, side-by-side tables with the IDE, charts and board packs

Study design

Balanced, seed 20260822:

Stratum	Levels	Per level
Market	US, ZA, CD, PH, SG, GB (England), FR, PL, IN, NG	1,000
Age band (attained age at 31 Dec 2025)	18–29, 30–39, 40–49, 50–59, 60–69, 70–84	1,660–1,670
MBTI type	all 16	625
Sex	M, F	5,000

Every market × age band × MBTI cell holds 10 or 11 agents. Within a market the socio-economic attributes (locale, community, income band, education, employment, marital status, children, mortgage, risk tolerance, financial literacy) are drawn from policyholder-population priors. The design is deliberately balanced, not census-weighted, so that every cell has exposure. Re-weight (post-stratify) with the attributes carried on each row before reading anything as a market statistic.

Personality enters as a plain-language money-behaviour profile for each MBTI type. The agent is told to let it colour how they handled the policy, not to override their circumstances.

How the records were made

1. Inversion. The request asks insurers for VM-51 records. Here each agent is one policyholder, answering a policyholder-facing questionnaire derived from Appendix B: the policy they hold, its amounts and underwriting, and its status in each of 2023, 2024 and 2025.
2. Conditioning. Each agent receives its persona, its own market's dated context block, and income and market anchors.
3. Generation. 10,000 calls to claude-sonnet-5 through Claude Code on the Scelo swarm's society tier, run on 22 August 2026. The raw replies are in raw/responses.jsonl.
4. Validation and repair. Every answer is checked against the product contract, and each of the 79 repairs is logged in policies.csv → repairs:
   • the issue date is at or after age 18 and no later than December 2025; statuses are forced to not_yet_issued before issue and terminated_earlier after the first termination;
   • term cannot be surrendered (except return-of-premium term), cannot go reduced paid-up or extended term, and can only expire once its level period has run;
   • whole life cannot convert or expire and has no cash value before year 2;
   • missing or implausible premiums are estimated and flagged premium_estimated (11 policies).
5. Mortality overlay. Agents decide only voluntary outcomes. Deaths are drawn by the driver from a Gompertz curve, q = 0.00004·e^(0.088·age) (female × 0.65, market factor 0.8–1.9), so the file carries cause-04 terminations and complete exposure. The overlay is imposed and carries no information.
6. Expansion. Each history becomes one VM-51 record per observation year, with central exposure.

VM-51 conventions and documented deviations

• Items 1–2 carry a pseudo submitting-company code, 999xx, one per market (there is no NAIC company code).
• Item 6 is the US state of issue for US policies and the ISO country code otherwise.
• Item 9 declares age last birthday.
• Amounts are converted to USD at the FX rates pinned in metadata/manifest.json; the local-currency amounts are kept as extension columns.
• Items 29–31 are filled only for level and return-of-premium term plan codes, as the specification requires.
• The UL/VUL items 33–46 are blank.

What came out (this describes the generator, not a market)

Policies	10,000
Policy-year records	27,692
Central exposure (policy-years)	25,775.5
Lapses	2,822
Surrenders	380
Deaths (overlay)	229
Reduced paid-up	182
Expiries	132
Conversions	1

Products: level term 3,727 · whole life non-participating 3,070 · whole life participating 2,676 · decreasing term 517 · annually renewable term 10.

Lapses per exposure-year, by market:

Market	Exposure	Lapses	Rate	95 % interval
Nigeria (NG)	2,167.9	525	0.242	0.222–0.264
Democratic Republic of the Congo (CD)	2,382.7	546	0.229	0.210–0.249
South Africa (ZA)	2,615.1	328	0.125	0.112–0.140
France (FR)	2,588.6	287	0.111	0.098–0.125
Philippines (PH)	2,658.1	292	0.110	0.098–0.123
Poland (PL)	2,615.4	230	0.088	0.077–0.100
England (GB)	2,678.6	212	0.079	0.069–0.091
United States (US)	2,692.3	163	0.061	0.052–0.071
India (IN)	2,588.2	151	0.058	0.049–0.068
Singapore (SG)	2,788.6	88	0.032	0.025–0.039
All markets	25,775.5	2,822	0.109

Two classic persistency shapes emerged without being imposed on the generator. Lapse falls monotonically with attained age, from 0.171 at 18–29 to 0.059 at 70–84. By policy duration it starts at 0.040 in year 1, peaks at about 0.15 in years 4 to 10, and falls to 0.024 after year 20. India combines one of the lowest lapse rates with by far the highest surrender rate (0.072 per exposure-year), which is what a savings-oriented market with cash values looks like.

The analysis: the term book on three engines

The term-life slice (term_book, 4,254 policies) was analysed three separate times: in Scelo IDE 0.1.6, with scelo for Python, and with scelo for R. The results were then reconciled. Every output is under analysis/.

Figure	Value	Where it agrees
Lapse GLM: Poisson, log link, offset log(exposure); country + product + MBTI	AIC 5,238.8 · deviance 2,388.8 · base rate 0.0775 (DRC · ART · ENFJ)	identical on all three engines
PV net cash flow, lifelib BasicTerm_ME, full term book (A)	1,017,152	IDE and scelo (lifelib 0.14.0)
PV net cash flow, lifelib BasicTerm_ME, in-force book (B)	1,098,674	IDE and scelo (lifelib 0.14.0)
PV net cash flow, scelo BasicTerm port, illustrative basis	980,637	Python and R
Solvency II life SCR, standard-formula shocks on the port, full book	1,260,819	Python and R
IFRS 17 CSM at initial recognition (RA 5 % of PV claims), full book	784,998	Python and R

The two libraries agree on 74 serialised figures to a maximum relative difference of 6.6 × 10⁻¹⁰. Against DRC, market relativities run from 0.19 (India) to 1.16 (Nigeria, not significant); product is not significant; and the MBTI relativities span 156×. That last spread is a designed property of the synthetic population, not an empirical finding.

Uses

Suitable for:

• validating experience-study ingestion, edit checks and aggregation pipelines end to end against the exact VM-51 shape;
• building and testing tooling and dashboards on a file that needs no clearance to share;
• teaching experience-study methods: exposure, occurrence–exposure rates, Poisson GLMs with an exposure offset, credibility, and IFRS 17 and Solvency II on a model-point file;
• research into language-model-generated synthetic populations and their biases.

Not suitable for:

• setting, validating or benchmarking any valuation, pricing or reserving assumption;
• drawing conclusions about any real market, insurer, product or population;
• drawing any conclusion about personality and insurance behaviour in real people.

Limitations and biases

• Synthetic. Every statistic describes the generator (a language model conditioned on personas and dated context), not a market.
• Designed personality signal. The MBTI relativities span 156×, far wider than in any real book. The signal was designed in.
• Acquiescence. 94.5 % of agents said they had considered lapsing. That is an artefact of asking the question, not a rate.
• Narrative-biased timing. Calendar-year timing within 2023–25 is narrative-biased. Estimate lapse as one force over the window with exposure as an offset, not as three calendar-year hazards.
• Imposed mortality. Deaths come from the Gompertz overlay and carry no information.
• Balanced, not census-weighted. Re-weight by post-stratification before reading anything as a market statistic.
• Model-written text and sensitive attributes. The free-text fields (why_bought, the yearly notes, voice) were written by the language model and may reproduce stereotypes about markets, communities or religions. Attributes such as community, sex and mbti are synthetic design variables: do not use them to infer anything about real groups, or as rating factors.
• No real people. Every person, policy and amount is synthetic. Any resemblance to a real person is coincidental.

Loading

from datasets import load_dataset

repo = "intelligentactuaries/synthetic-lapse-surrender-study-2023-2025"
records = load_dataset(repo, split="train")                # VM-51 policy-years (default config)
policies = load_dataset(repo, "policies", split="train")

import pandas as pd

base = "hf://datasets/intelligentactuaries/synthetic-lapse-surrender-study-2023-2025"
df = pd.read_csv(f"{base}/data/vm51_policy_years.csv")
replies = pd.read_json(f"{base}/raw/responses.jsonl", lines=True)   # the raw model replies

With scelo (pip install "scelo[stats]"), reproduces the published lapse GLM:

import scelo as sc
from huggingface_hub import hf_hub_download

path = hf_hub_download("intelligentactuaries/synthetic-lapse-surrender-study-2023-2025",
                       "data/term_book_2023_2025.csv", repo_type="dataset")
book = sc.load(path)
book["lapsed"] = (book.final_status == "lapsed").astype(int)
glm = sc.glm(book, "lapsed ~ C(country) + C(product) + C(mbti)", "poisson",
             offset="exposure", base="first")
glm.relativities()   # AIC 5,238.8 · base rate 0.0775

Citation

@misc{denewade2026syntheticlapse,
  author       = {Denewade, Ali},
  title        = {Synthetic Term and Whole Life Lapse and Surrender Study, 2023--2025:
                  10,000 synthetic policyholders in the VM-51 layout},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/intelligentactuaries/synthetic-lapse-surrender-study-2023-2025}},
  note         = {Intelligent Actuaries. Synthetic data, generated in response to the SOA Research
                  Institute and LIMRA 2023--2025 Term and Whole Life Lapse and Surrender Experience
                  Study data request. Not SOA or LIMRA data.}
}

Please also cite the request it answers:

@misc{soalimra2026request,
  author       = {{Society of Actuaries Research Institute} and {LIMRA}},
  title        = {2023--2025 Term and Whole Life Lapse and Surrender Experience Study Data Request},
  year         = {2026},
  month        = jul,
  howpublished = {\url{https://www.soa.org/research/opportunities/2026/2023-2025-term-whole-life-lapse-sur-exp-study-data-request/}}
}

Licence

The data is released under CC BY 4.0. The SOA/LIMRA data-request document is not included and remains © 2026 Society of Actuaries Research Institute and LIMRA.

Contact

Intelligent Actuaries · research@intelligentactuaries.com · intelligentactuaries.com