Mortality Forecasting as a Flow Field in Tucker Decomposition Space

Authors: Samuel J. Clark Date: 2026-03-25 Paper ID: openalex:2603.24299

Summary

This paper reframes mortality forecasting by modeling the evolution of multi-population mortality schedules as the integration of a continuous flow field within the low-dimensional latent space defined by a Tucker tensor decomposition of the data. The analysis reveals that the primary dynamic is a one-dimensional flow governed by a scalar speed function, while structural scores provide the remaining dimensions. The proposed system employs an era-weighted speed function to adapt to current dynamics and uses empirically calibrated convergence rates to ensure relaxation toward canonical mortality structures. Performance is rigorously tested via leave-country-out cross-validation over a 50-year horizon, benchmarking against established methods like Lee-Carter.

Key Contributions

• Reframing mortality forecasting as integrating a flow field within the latent score space of a Tucker tensor decomposition of multi-population data.
• Demonstrating that mortality transition in this latent space is largely a one-dimensional flow driven by a scalar speed function controlling the life expectancy level.
• Developing an era-weighted speed function to adapt the flow dynamics to contemporary forecasting origins.
• Evaluating the system using leave-country-out cross-validation with a 50-year horizon against Lee-Carter and Hyndman-Ullah benchmarks.

Limitations

The paper’s abstract does not explicitly state limitations, but the complexity of calibrating the convergence rates and the assumption of a dominant one-dimensional flow are potential areas for further investigation.

Key Concepts

• tucker-decomposition-mortality-forecasting: A method to model mortality forecasting as integrating a flow field within the latent score space derived from a Tucker tensor decomposition of population mortality data.
• mortality-flow-field-integration: Modeling the evolution of mortality structure as the integration of a continuous velocity field within the latent space of a Tucker decomposition.

Archivist Review

Two central concepts were approved: the overall framework of using Tucker Decomposition for this specific task, and the core mechanism of modeling the evolution as a ‘Flow Field Integration’ in the latent space, as both represent reusable conceptual shifts in time-series modeling. The named dataset, Human Mortality Database, was approved as a critical benchmark source. No open questions were deemed substantial enough to warrant a standalone entry beyond general research directions.

Approved Concepts

• Tucker Decomposition for Mortality Forecasting: This method reframes the entire forecasting problem as a continuous flow in a low-dimensional latent space derived from Tucker decomposition, which is a novel structural approach to mortality modeling.
• Mortality Flow Field Integration: This describes the core dynamic mechanism: modeling the evolution of mortality structure not as discrete steps but as continuous movement (flow) along latent structural axes.

Rejected Candidates

• [concept] Tucker Decomposition for Mortality Forecasting (tucker-decomposition-mortality-forecasting) - duplicate_existing: The existing concept ‘tucker-decomposition-mortality-forecasting’ is already approved. A new candidate with the same slug is not needed.
• [concept] Mortality Flow Field Integration (mortality-flow-field-integration) - duplicate_existing: The existing concept ‘mortality-flow-field-integration’ is already approved. A new candidate with the same slug is not needed.

Datasets

• human-mortality-database

Links

• Abstract
• PDF