Bell 1997 — Comparing and Assessing Time Series Methods for Forecasting Age-Specific Fertility and Mortality Rates

Summary

Bell (1997) conducts an out-of-sample forecast accuracy evaluation comparing four time-series methods for age-specific fertility and mortality rates: Heligman-Pollard parametric curve fitting (HP), the Lee-Carter single-principal-component model (LC), the Bell-Monsell full-principal-component model (BM), and age-specific random walk with drift (RWD). The central finding is that LC, HP, and BM all fail to beat the atheoretic RWD benchmark when implemented without a bias adjustment — but LC with bias adjustment (initializing forecasts from the most recent observed rates rather than the fitted age averages) performs approximately equivalently to RWD. This motivates the now-standard "jump-off fix" for Lee-Carter implementations and establishes bias adjustment as a non-negotiable requirement for any structured curve-fitting or low-dimensional principal-component (PC) method.

Key Claims

• Four methods compared: HP (Heligman-Pollard $8$-parameter mortality curve fit to each year, extrapolated), LC (Lee-Carter singular value decomposition (SVD), 1 PC with $k_t$ as random walk with drift), BM (Bell-Monsell full-PC SVD decomposition), RWD (independent random walk with drift per age group, no age-schedule structure imposed)
• RWD baseline: mean absolute percentage error (MAPE) of approximately $2.6$–$6.1\%$ across rate types and forecast horizons; serves as the atheoretic benchmark that all structured methods must beat to justify their complexity
• LC without bias adjustment loses to RWD: MAPE ratio LC/RWD $\approx 2$–$3\times$; the structured SVD decomposition fails to improve on age-specific random walks when the forecast is initialized from the smoothed age-average profile $a_x$ rather than the most recent observed rates
• LC with bias adjustment $\approx$ RWD: After initializing from most recent observed rates (jump-off fix), MAPE ratio $\approx 0.98$–$1.07\times$ — near-equivalence; the structured model adds no loss and no gain over the baseline
• HP without bias adjustment: MAPE ratio HP/RWD $\approx 1.3$–$2.5\times$; curve-fitting induces a systematic jump-off bias similar to LC but also suffers from parametric misspecification at extreme ages
• Bell-Monsell (all PCs): MAPE ratio BM/RWD $\approx 1.03$–$1.22\times$; using all principal components rather than just the first captures more of the short-run age dynamics, so BM performs better than LC-without-bias even without a bias adjustment — but still marginally worse than RWD
• Firm conclusion: Bias adjustment is essential for any structured curve-fitting or low-dimensional PC method; without it, the age-specific random walk dominates uniformly across rate types and horizons
• Tentative conclusion: Even with bias adjustment, structured methods offer no systematic short-to-medium-horizon accuracy advantage over age-specific random walks — the parsimony of RWD is sufficient for forecasting accuracy even if structured models offer other benefits (uncertainty quantification, interpretability)

Concepts Introduced or Extended

• Lee-Carter Model — jump-off fix motivated and quantitatively validated; bias adjustment shown to be the dominant driver of LC forecast accuracy
• Stochastic Fertility Forecasting — HP and RWD evaluated on fertility as well as mortality; bias adjustment finding from Thompson et al. 1989 — Multivariate Time Series Projections of Parameterized Age-Specific Fertility Rates confirmed as a general property, not fertility-specific

Entities Mentioned

• William R. Bell
• Ronald Lee
• Lawrence Carter

My Take

The paper's lasting methodological contribution is the empirical grounding of the jump-off fix: what had been an ad hoc implementation choice (initialize from recent observed rates rather than the smoothed $a_x$) is now a validated necessity — without it, the entire framework loses to a naive benchmark. The RWD equivalence result is methodologically important but should not be over-read: RWD and LC+bias produce similar MAPE in short-to-medium horizons, but only LC provides a structured framework for uncertainty quantification via the $k_t$ autoregressive integrated moving average (ARIMA) variance. The Bell-Monsell finding (all-PC model beats 1-PC model without bias) anticipates later work showing that the first PC captures most but not all of the mortality co-movement structure; subsequent authors (Meseguer 2006, Pedroza-King 2002) moved to Bayesian state-space representations that address this more formally. The paper's cautionary lesson — that model parsimony plus the right initialization beats model complexity with the wrong initialization — generalizes well beyond mortality forecasting.