Actuarial-Informed Neural Networks: Constrained Deep Learning for Sex-Specific Multi-Population Longevity Forecasting
Authors: Davide Rindori
Organizations: Centre of Excellence for Global Finance Operations and Transformation Swiss Re Management Ltd∗ Mythenquai 50/60, 8022 Zurich, Switzerland
We introduce Actuarial-Informed Neural Networks (AINN), a constrained deep learning framework for sex-specific multi-population longevity forecasting that embeds actuarial knowledge into LSTM training through differentiable penalty terms. Applied to a high-longevity 6-country cluster (Switzerland, Sweden, Norway, West Germany, Netherlands, Japan; 1956-2020), the framework is a challenger to the Li-Lee benchmark and unconstrained neural approaches. Our central finding is that monotonicity constraint intensity lambda functions as a continuous governance dial with a quantifiable accuracy-stability trade-off. At lambda = 0.01, 30-year forecast confidence intervals narrow by 6.5% at zero RMSE cost. At lambda = 0.1, they narrow by 25% at only +0.84% RMSE cost. This trade-off is invisible to standard hyperparameter optimizers, which select lambda by minimizing one-step RMSE, the wrong metric for structural constraints in recursive forecasting. The framework adds three contributions: (i) residual-calibrated process noise that eliminates double-counting in the dual uncertainty framework, reducing confidence intervals by 53-60%; (ii) a 5-seed model ensemble with a coefficient of variation of 1.06%; and (iii) a time-varying credibility weight Z(t) = exp(-alpha t) connecting to Buhlmann credibility theory for blending neural and actuarial predictions. A 10-year out-of-sample backtest (trained on 1956-2010, validated against observed HMD life expectancy for 2011-2020) yields a mean absolute error of 0.38 years, with 11 of 12 country x sex combinations within the 90% confidence interval. The model produces regulatory-grade outputs including Solvency Capital Requirements under the Swiss Solvency Test (ES 99.0%: +1.91 years, CHE Male) and Solvency II, reverse stress testing, and model-based stability verification (amplification ratio 1.02x).
Figures & tables
Figure 1: Log-mortality at age 65 by sex across the frontier cluster (1956–2020). Male mortality (left) shows steeper improvement and higher cross-country volatility than Female mortality (right), reflecting the well-documented “male mortality convergence” phenomenon driven by reductions in cardiovascular disease and smoking-related deaths.
Figure 2: Li-Lee common mortality index Kt,s by sex. The Male trend declines more steeply, reflecting faster historical mortality improvement. The post-2000 convergence of the two trends is consistent with the closing of the sex-mortality differential.
Country
Combined [ 9 ]
Male
Female
Switzerland
Conflict
Stationary
Conflict
Sweden
Unit Root
Stationary
Stationary
Norway
Stationary
Conflict
Conflict
West Germany
Unit Root
Unit Root
Unit Root
Netherlands
Unit Root
Unit Root
Unit Root
Japan
Conflict
Conflict
Conflict
Table 1: Stationarity Analysis: Combined [ 9 ] vs. Sex-Specific
Parameter
Values
Rationale
Lookback L
{5, 7, 10, 12, 15}
Temporal context
Units u1
{16, 32, 48, 64}
First LSTM capacity
Units u2
{8, 16, 24, 32}
Second LSTM capacity
Learning rate
{0.01, 0.005, 0.001, 0.0005, 0.0001}
Optimizer step
λcoherence
{0, 0.001, 0.01, 0.1, 1.0}
Coherence weight
λmonotonicity
{0, 0.001, 0.01, 0.1, 1.0}
Monotonicity weight
Table 2: Hyperparameter Space for Joint Optimization
Figure 3: Constraint intensity trade-off across four metrics. The axes represent distinct concerns: one-step accuracy (top-left), biological compliance (top-right), and forecast stability for Male and Female (bottom). The “sweet spot” at λ=0.1 reduces CI by 25% at only +0.84% RMSE cost.
Figure 4: Stochastic Kt fan charts for Male (left) and Female (right), 2020–2050. The 80% and 95% confidence intervals are derived from the 5-seed ensemble with residual-calibrated process noise. The Female fan chart is narrower despite a historically larger σ , reflecting the greater σ reduction from residual calibration ( −60% vs. −53% ).
Figure 5: Life expectancy convergence across the cluster by sex. West Germany exhibits the steepest catch-up trajectory for both sexes, consistent with the persistent unit root in its country-specific factors. Japan Female has the smallest gain (+1.64 years) because it starts from the highest baseline (84.89), approaching a plausible biological ceiling.
Sex
Model
e0 (2050)
Gain
CI Width
SCR (ES)
M
Li-Lee
83.54
+3.28
3.73
+2.091
M
AINN
82.06
+1.80
3.69
+1.907
F
Li-Lee
86.05
+2.47
4.71
+2.178
F
AINN
85.55
+1.96
3.03
+1.516
Table 8: AINN Ensemble vs. Li-Lee RWD — Switzerland at 2050
Figure 6: Switzerland e0 : AINN ensemble (colored, with CI) vs. Li-Lee RWD (grey). The AINN captures the post-2011 deceleration, producing more conservative median projections. The advantage is most visible for females, where the AINN’s residual-calibrated CI is 36% narrower than Li-Lee’s.
Country
Sex
Observed
AINN
Li-Lee
AINN Error
In 90% CI
CHE
M
80.26
80.52
80.51
+ 0.26
YES
CHE
F
83.58
83.84
84.04
+ 0.26
YES
NOR
M
80.58
79.56
79.54
− 1.02
NO
JPN
F
84.89
84.78
84.95
− 0.11
YES
Table 9: 10-Year Backtest: Predicted vs. Observed e0 at 2020 (Selected Countries)
Figure 7: 10-year backtest for Switzerland: predicted vs. observed e0 (red circles). Both models track the trajectory reasonably well; the AINN’s advantage is in uncertainty calibration (narrower CI for females) rather than in point accuracy.
Figure 8: Temporal saliency profiles for Male and Female. The distributed importance across the full 15-year window contrasts with the concentrated t−4 peak (21.7%) observed in [ 9 ] with L=10 , suggesting that the longer window allows the model to integrate evidence more broadly.
Figure 9: SHAP influence ranking for Swiss Male mortality. The distributed pattern across all cluster members indicates that the LSTM integrates the full multi-population structure. The common factor Kt ranks lowest because its marginal contribution over the background is minimal (all samples share a similar Kt pattern).
Figure 10: Lexis residual maps for Switzerland (ages 40–90). The smooth vertical gradient with no diagonal cohort artefacts confirms that the rank-1 reconstruction captures the dominant age-period interaction without introducing spurious cohort effects.
Country
Male SCR
Female SCR
Switzerland
+1.907
+1.516
Sweden
+1.942
+1.649
Norway
+1.852
+1.580
W. Germany
+2.245
+1.800
Netherlands
+2.063
+1.849
Japan
+1.902
+1.243
Table 10: SCR (ES 99.0%, SST) — Full Cluster
Figure 11: Tail risk distribution for Swiss e0 at 2050 (Male and Female). The ES 99.0% thresholds are indicated, defining the SCR under the SST framework.
Figure 12: Model-based stress test response. A negative shock of −2σ in ΔKt space is injected at 2026. The model absorbs the shock without amplification (maximum ratio 1.02 × ), confirming numerical stability under perturbation.
α
Z (2025)
Z (2035)
Z (2050)
M Med
F Med
M CI
0.0
1.000
1.000
1.000
82.06
85.55
3.69
0.02
0.905
0.741
0.549
81.69
85.14
2.89
0.05
0.779
0.472
0.223
81.26
84.73
2.16
0.10
0.607
0.223
0.050
80.90
84.33
1.81
0.20
0.368
0.050
0.002
80.63
83.98
1.35
Table 11: Credibility Z(t)=e−αt : Effect on CHE Projections
School of Computing and Augmented Intelligence, Arizona State University, Tempe, USA · School of Life Sciences, Arizona State University, Tempe, USA · School of Mathematical and Statistical Sciences, Arizona State University, Tempe, USA