Evaluating the performance of the Lee-Carter method for forecasting mortality

Area of Science:

• Demography
• Actuarial Science
• Biostatistics

Background:

• The Lee-Carter (LC) model, a prominent statistical method for mortality forecasting introduced in 1992, is evaluated for its predictive accuracy.
• This study assesses both historical and simulated forecast errors of the LC method, comparing them against errors from Social Security Administration projections.

Discussion:

• The LC method generally underprojected mortality improvements, but to a lesser extent than Social Security projections.
• While ex ante 95% probability intervals contained the true life expectancy (e0) 97% of the time overall, these intervals exhibited biases: too wide for shorter horizons (up to 40 years) and too narrow for longer horizons (after 50 years).

Key Insights:

• Projections made post-1945 for the period up to 1998 consistently showed forecast errors under two years.
• Hypothetical applications of the LC method to mortality data from France, Sweden, Japan, and Canada indicated strong performance in these countries.

Outlook:

• The effectiveness of the LC method may be challenged by evolving age-specific patterns in mortality decline throughout the 20th century.
• Further research could explore model adjustments to better accommodate dynamic shifts in mortality trends across different age groups and populations.