Neural Mortality Benchmark

When should actuaries trust neural networks for mortality forecasting?

A reproducible benchmark of 14 classical, baseline, neural and hybrid mortality-forecasting implementations. The study uses rolling-origin validation, actuarial error metrics and an illustrative French annuity case study to test when additional model complexity is useful.

Research status: independent working paper, version 0.1.0. The manuscript has not been peer reviewed and is not presented as an accepted publication.

Author

Amine Manai — M1 Actuarial Science at Le Mans Université, Institut du Risque et de l’Assurance; double-degree student in Data Science Engineering at ESPRIT.

• Working paper (PDF)
• French technical report
• Methodology and defense notes
• Stored benchmark results

Study design

• 14 implementations: five classical structural models, two simple baselines, six neural architectures and the proposed LC-ResNet hybrid.
• Eight national populations: France, England and Wales, USA, Japan, Italy, Spain, Sweden and the Netherlands.
• Rolling-origin validation: origins from 1990 to 2018 and horizons of 1, 5, 10 and 20 years when observations are available.
• Actuarial metrics: error on period life expectancy at birth and an annuity-due factor at age 65, both truncated at age 100.
• Targeted scenarios: short histories, the reported 2020–2022 mortality-shock window and separate age segments.
• Illustrative liability study: 1,000 French annuitants valued with 2033 projected rates and a permanent 20% mortality-rate reduction.

Classical and baseline	Neural	Hybrid
Lee-Carter, Lee-Miller, Poisson LC, CBD, H-U-style functional model, age-specific random walk with drift, frozen rates	LSTM, GRU, Bi-LSTM and Transformer on κₜ; FFNN with embeddings; mortality-surface CNN	LC-ResNet

The H-U implementation is deliberately simplified: spline smoothing, six functional components and random-walk-with-drift score forecasts. It is not the full robust automatic-ARIMA procedure.

Main findings

• The age-specific random walk with drift has the lowest full-grid log-rate RMSE at every tested horizon.
• LC-ResNet has the lowest one-year error among structured models; Lee-Miller performs better at 20 years.
• The H-U-style model has the lowest 20-year life-expectancy RMSE, while the FFNN has the lowest 20-year annuity-factor RMSE.
• Frozen rates lead the cumulative 2020–2022 stress test and the 20-year-history scenario.
• In the illustrative French case study, the five-model base-liability spread is €6.26 million; it is €1.22 million among the four full-age-grid models.
• The liability increase under the 20% longevity stress varies from €11.52 million to €12.33 million across the five models.

These findings are specific to the selected populations, period, origins, loss functions and training budget. They are evidence from this benchmark, not universal model rankings.

LC-ResNet

LC-ResNet combines:

1. a Poisson Lee-Carter skeleton (aₓ, bₓ, κₜ);
2. a small neural network trained on the structured residuals;
3. horizon-dependent shrinkage, where the neural correction is multiplied by exp(−λh).

The Lee-Carter component remains reportable, but the residual network is opaque. The complete hybrid is therefore partially interpretable, not fully interpretable.

Reproduce the project

1. Install the locked environment

git clone https://github.com/aminemanai2003/neural-mortality-benchmark.git
cd neural-mortality-benchmark
uv sync --all-extras --frozen

Python 3.11 or later and uv are required. A hash-locked pip export is also available in requirements.lock.

2. Download HMD data

Create a local .env from .env.example, then add credentials for your free HMD account:

HMD_USERNAME=your-email@example.com
HMD_PASSWORD=your-password

uv run python scripts/download_hmd.py

The downloader writes an untracked manifest containing retrieval timestamps and file hashes for future runs.

3. Run the analyses

# Reduced benchmark for a fast functional check
uv run python scripts/run_benchmark.py --countries FRATNP --quick

# Complete benchmark and the French case study
uv run python scripts/run_benchmark.py
uv run python scripts/run_case_study.py

# Dashboard, tests and summary chart
uv run streamlit run dashboard/app.py
uv run pytest
uv run python scripts/plot_summary.py

Data licence and reproducibility boundary

The project uses Human Mortality Database estimates. HMD-produced estimates are distributed under CC BY 4.0, while source input data may remain subject to the original providers’ terms. HMD recommends that users download a current copy rather than pass around local copies. See the HMD User Agreement and citation guidance.

Raw data are therefore not committed here. The repository can re-execute the study for an authorized HMD user, but the exact download date of the snapshot behind the checked-in 2026 results was not recorded. HMD revisions may cause future reruns to differ slightly; new downloads now create a local manifest to preserve that provenance.

Repository map

src/mortality/       data loading, models, evaluation and actuarial calculations
scripts/             benchmark, case study, download and chart entry points
config/              data, model and evaluation settings
tests/               automated unit and integration tests
results/             stored benchmark rows, case study and summary figure
paper/               LaTeX working-paper source and compiled PDF
report/              French technical report and generated DOCX
dashboard/           Streamlit application

Important limitations

• National-population mortality is not insured-portfolio mortality.
• The study uses one neural seed and no model-specific hyperparameter search.
• Forecasts are point estimates; interval calibration and formal pairwise tests are not reported.
• Actuarial metrics stop at age 100.
• Countries contribute different numbers of valid origins because their series end in different years.
• The 20% longevity stress is applied to liabilities only; its output is not a complete insurer-level SCR calculation.
• The precise HMD retrieval date for the stored benchmark was not preserved.

Citation and licence

Citation metadata are provided in CITATION.cff. The software is released under the MIT License. HMD data remain governed by their own licence and source-provider terms.