pyliferisk geometric valuations discard supplied mortality assumptions

English evidence video

Watch the published YouTube Short · LinkedIn post · Original evidence ZIP · Video source package

46 seconds. Original diagrams and synthetic Microsoft Andrew English narration via the open-source edge-tts client and Microsoft online synthesis. Alex Vector is a fictional narrator. Complete normal-speed playback and technical checks passed. The video explains the recorded synthetic cases; no insurer exposure or customer loss is asserted.

GitHub report and source · Hugging Face · Zenodo

Archival mirror. Original report. Claims, dates, authorship and licenses remain those of the original publication; this catalog update does not rerun or revalidate its numerical experiments.

Xamit Kadirbekov · GERO Research · 14 September 2026

Status: reproduced on the official PyPI 1.12.0 wheel; correction submitted as upstream PR #17. The patch is proposed, not an accepted or released upstream fix.

A zero-growth annuity changes value

An annuity with zero growth should equal the corresponding level annuity when both use the same interest rate and mortality assumptions. In the released library, qax() instead resets mortality to the original table's 100% level.

from pyliferisk import Actuarial, ax, qax
from pyliferisk.mortalitytables import GKM95

mt = Actuarial(nt=GKM95, perc=75, i=0.03)
print(ax(mt, 60))       # 14.93597449181164
print(qax(mt, 60, 0))   # released: 13.616328838004055

This is a synthetic annual whole-life annuity-immediate of one unit, age 60, interest 3%, and mortality rates scaled to 75% of GKM95. A direct conditional survival cash-flow sum returns 14.935974491811635. The released geometric result is approximately 8.8353% lower. This is a library pricing discrepancy, not a claim of actual insurer exposure or customer losses.

Mortality scale	Level annuity / direct cash-flow value	Released zero-growth qax
50%	16.80035314426194	13.616328838004055
75%	14.93597449181164	13.616328838004055
100% control	13.616328838004055	13.616328838004055

The same cause affects qAx, qax, qaax, qaxn, qaaxn, qtax, and qtaax. Public annuity() dispatch reaches affected helpers. Tables supplied directly as survivor counts (lx) or mortality probabilities (qx) have no original nt; geometric revaluation can then raise an exception or use an empty table.

Cause and proposed correction

The helpers adjust interest to j = (i - growth) / (1 + growth) and construct Actuarial(nt=mt.nt, i=j). That changes both the discount rate and the mortality table: the constructor defaults perc to 100 and cannot recover custom lx or qx from nt=None.

The correction supplies copies of the effective mortality data instead:

mtj = Actuarial(lx=mt.lx[:], qx=mt.qx[:], i=j)

Only the discount rate changes. Copies prevent constructor writes from aliasing the caller's lists. The patch preserves existing growth timing and fractional payment approximations. It does not change the separate undefined-variable paths qAxn/qtAx, unfinished helpers, deferred geometric dispatch, or unrelated fractional-payment issues. Those are not claimed as corrected here.

Verification

• 630 annual scenarios, using independent direct sums at 60-digit Decimal precision: 498 failures before, zero after. The 498 include 198 exceptions; exceptions are not an additional count.
• GKM95 at 50%, 75%, and 100%; custom survivor and probability tables; interest rates -1%, 0%, and 3%; growth -2%, 0%, and 2%; two ages and seven helpers.
• Four standard-library regression test methods pass. They also check zero growth with annual, semiannual, quarterly and monthly payments, public non-deferred annuity() dispatch, and unchanged caller state.
• Running the new tests against the original released module produces 82 failing and 42 errored subcases. These are subcases, not 124 test methods.
• The Decimal oracle uses the supplied survivor table and sums each actual conditional payment directly; it does not use commutation functions or an Actuarial reconstruction at the transformed rate. Agreement tolerance is rel_tol=2e-12, abs_tol=2e-12.

All calculations ran sequentially on CPU with Python 3.9.6 and the standard library. No GPU or parallel worker was used. There is no performance benchmark. The suite establishes these cases, not correctness of the entire library.

Reproduce

The official wheel is included for reproducibility under its GPL license. The vendor/released and vendor/patched directories contain the relevant source snapshots. From this directory:

python3 -B verify.py --source vendor/released --output released-rerun.json
python3 -B verify.py --source vendor/patched --output patched-rerun.json
PYTHONPATH=vendor/patched python3 -B -m unittest discover -s tests -v

The verifier records mismatches in JSON; it exits normally for an expected failing baseline, so inspect the failures and exceptions fields.

Official wheel pyliferisk-1.12.0-py3-none-any.whl SHA-256: 1c2ab2b902424f33d408d5f4e13badd98d3d0ca26462d83562bb6b21cd2bca53. The released __init__.py is byte-identical to upstream master 5c28ca34a30f4e350f604f4b35bf738c816e59bc (17 September 2023). Correction commit: 0458d03ca4d03a73249df3e70c033167235962fa.

Duplicate screening and sources

On 14 September 2026, screening covered all 14 available upstream issue/PR records and all 8 available issue comments before submission. GitHub issue searches scoped to the repository for geometric, perc, and qax each returned zero matches. The earlier reports #15 (deferred mortality) and #16 (deferred fractional-payment adjustment) concern different expressions. No matching earlier report was found in that scope. This is not a guarantee of worldwide novelty or a claim that no private report exists.

• Official PyPI release and artifact hashes
• Pinned upstream source
• Project documentation of the perc parameter
• Proposed upstream correction and review status

Original code and mortality tables: Francisco Garate and contributors, GPL-3.0-or-later. This report, regression tests, and reproduction scripts are also provided under GPL-3.0-or-later. See LICENSE.

Publication record

This page was added on 14 September 2026 to align the public archive. The original report, source history and licensing remain available from the links above. AI-assisted archival preparation.

Complete cross-platform publication record